Basic Filters
2026 |
Andreas Kontogiannis Vasilis Pollatos, Panayotis Mertikopoulos Ioannis Panageas Efficient swap regret minimization in combinatorial bandits Conference Twenty-Ninth Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026), 2026. @conference{Kontogiannis2026, title = {Efficient swap regret minimization in combinatorial bandits}, author = {Andreas Kontogiannis, Vasilis Pollatos, Panayotis Mertikopoulos, Ioannis Panageas}, url = {https://arxiv.org/pdf/2602.02087}, year = {2026}, date = {2026-05-02}, booktitle = {Twenty-Ninth Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026)}, abstract = {This paper addresses the problem of designing efficient no-swap regret algorithms for combinatorial bandits, where the number of actions N is exponentially large in the dimensionality of the problem. In this setting, designing efficient no-swap regret translates to sublinear — in horizon T — swap regret with polylogarithmic dependence on N. In contrast to the weaker notion of external regret minimization – a problem which is fairly well understood in the literature – achieving no-swap regret with a polylogarithmic dependence on N has remained elusive in combinatorial bandits. Our paper resolves this challenge, by introducing a no-swap-regret learning algorithm with regret that scales polylogarithmically in N and is tight for the class of combinatorial bandits. To ground our results, we also demonstrate how to implement the proposed algorithm efficiently — that is, with a per-iteration complexity that also scales polylogarithmically in N — across a wide range of well-studied applications.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } This paper addresses the problem of designing efficient no-swap regret algorithms for combinatorial bandits, where the number of actions N is exponentially large in the dimensionality of the problem. In this setting, designing efficient no-swap regret translates to sublinear — in horizon T — swap regret with polylogarithmic dependence on N. In contrast to the weaker notion of external regret minimization – a problem which is fairly well understood in the literature – achieving no-swap regret with a polylogarithmic dependence on N has remained elusive in combinatorial bandits. Our paper resolves this challenge, by introducing a no-swap-regret learning algorithm with regret that scales polylogarithmically in N and is tight for the class of combinatorial bandits. To ground our results, we also demonstrate how to implement the proposed algorithm efficiently — that is, with a per-iteration complexity that also scales polylogarithmically in N — across a wide range of well-studied applications. |
Georgios Bouchouras Dimitrios Doumanas, Andreas Soularidis Konstantinos Kotis George Vouros Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting Journal Article AI, 7 (4), pp. 139, 2026, ISSN: 2673-2688. @article{Bouchouras2026, title = {Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting}, author = {Georgios Bouchouras, Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis, George Vouros}, url = {https://www.mdpi.com/2673-2688/7/4/139}, doi = {https://doi.org/10.3390/ai7040139}, issn = {2673-2688}, year = {2026}, date = {2026-04-14}, journal = {AI}, volume = {7}, number = {4}, pages = {139}, abstract = {Ontology engineering plays a critical role in clinical decision support systems for Parkinson’s Disease (PD) monitoring and alerting. While Large Language Models (LLMs) have shown promise in knowledge modeling tasks, their effectiveness in autonomously constructing comprehensive ontologies for complex clinical domains remains unclear. This study investigates four ontology engineering methodologies for PD monitoring and alerting: One-shot (OS) prompting, Decomposed Sequential Prompting (DSP), X-HCOME, and SimX-HCOME+. Multiple LLMs were evaluated across these methodologies. Generated ontologies were assessed against a reference PD ontology using structural evaluation metrics focused on classes and object properties. Expert review was additionally conducted to analyze knowledge extensions beyond the gold standard. LLMs were able to autonomously generate syntactically valid and semantically meaningful ontologies using OS and DSP prompting; however, these ontologies exhibited limited conceptual coverage. Incorporating human expertise through X-HCOME significantly improved ontology completeness and evaluation metrics. Expert review further validated clinically relevant concepts absent from the reference ontology. SimX-HCOME+ demonstrated that iterative, supervised collaboration supports ontology refinement, although challenges persisted in natural language-to-rule formalization. The findings suggest that LLMs are more effective as collaborative assistants rather than standalone ontology engineers in the PD domain. Structured human–LLM collaboration is associated with improved ontology coverage and facilitates the identification of potential knowledge extensions in clinical monitoring applications. While the present evaluation focuses primarily on structural ontology elements, the proposed methodologies provide useful insights for LLM-assisted ontology engineering in complex healthcare domains.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Ontology engineering plays a critical role in clinical decision support systems for Parkinson’s Disease (PD) monitoring and alerting. While Large Language Models (LLMs) have shown promise in knowledge modeling tasks, their effectiveness in autonomously constructing comprehensive ontologies for complex clinical domains remains unclear. This study investigates four ontology engineering methodologies for PD monitoring and alerting: One-shot (OS) prompting, Decomposed Sequential Prompting (DSP), X-HCOME, and SimX-HCOME+. Multiple LLMs were evaluated across these methodologies. Generated ontologies were assessed against a reference PD ontology using structural evaluation metrics focused on classes and object properties. Expert review was additionally conducted to analyze knowledge extensions beyond the gold standard. LLMs were able to autonomously generate syntactically valid and semantically meaningful ontologies using OS and DSP prompting; however, these ontologies exhibited limited conceptual coverage. Incorporating human expertise through X-HCOME significantly improved ontology completeness and evaluation metrics. Expert review further validated clinically relevant concepts absent from the reference ontology. SimX-HCOME+ demonstrated that iterative, supervised collaboration supports ontology refinement, although challenges persisted in natural language-to-rule formalization. The findings suggest that LLMs are more effective as collaborative assistants rather than standalone ontology engineers in the PD domain. Structured human–LLM collaboration is associated with improved ontology coverage and facilitates the identification of potential knowledge extensions in clinical monitoring applications. While the present evaluation focuses primarily on structural ontology elements, the proposed methodologies provide useful insights for LLM-assisted ontology engineering in complex healthcare domains. |
Dimitrios Doumanas Andreas Soularidis, Nikolaos Zafeiropoulos Stamatis Chatzistamatis George Tsekouras Andreas El Saer Chrisaphis Nathanailidis Konstantinos Kotis E Unbiasing Greek: In-Context Learning Strategies for Gender Bias Identification and Mitigation for Legal Documents and Job Ads Journal Article Information, 17 (4), pp. 342, 2026. @article{Doumanas2026b, title = {Unbiasing Greek: In-Context Learning Strategies for Gender Bias Identification and Mitigation for Legal Documents and Job Ads}, author = {Dimitrios Doumanas, Andreas Soularidis, Nikolaos Zafeiropoulos, Stamatis Chatzistamatis, George E Tsekouras, Andreas El Saer, Chrisaphis Nathanailidis, Konstantinos Kotis}, url = {https://www.mdpi.com/2078-2489/17/4/342}, doi = {https://doi.org/10.3390/info17040342}, year = {2026}, date = {2026-04-02}, journal = {Information}, volume = {17}, number = {4}, pages = {342}, abstract = {Gender bias embedded in legal and professional texts perpetuates systemic inequality, yet research on bias identification and mitigation remains largely confined to English. Morphologically rich languages such as Greek, where grammatical gender pervades nouns, adjectives, pronouns, and participles, present unique challenges that existing approaches fail to address. This paper elaborates on a systematic methodology primarily focusing on identifying and mitigating gender bias in Greek-language job advertisements and legal documents. To accomplish that task, we define a taxonomy of nine gender bias rules tailored to the linguistic properties of Greek and construct domain-specific annotated datasets comprising 90 expert-curated few-shot examples across both textual domains. Using these resources, we employ XML-structured prompt engineering with in-context learning (ICL)and systematically compare three classes of models: (i) commercial large language models (LLMs), namely Claude Sonnet 4.5 and GPT-5.2, (ii) two open-weight small language models (SLMs), Mistral Small (24B) and Ministral (14B), and (iii) Llama Krikri (8B), a Greek-native language model built on Llama 3.1 and fine-tuned on high-quality Greek corpora. For each input text, the system identifies biased expressions, maps them to specific bias rules, provides explanations, and generates a fully corrected inclusive version. Our experiments reveal substantial performance disparities across model scales and linguistic specialization, with LLMs demonstrating superior contextual reasoning and SLMs exhibiting systematic over-correction and grammatical errors in Greek morphology. We further introduce a critical meta-rule addressing gender agreement with named entities to prevent spurious corrections in legal texts referencing identified individuals. The findings highlight the importance of model scale, language-specific adaptation, and carefully designed prompting strategies for bias mitigation in underrepresented languages.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Gender bias embedded in legal and professional texts perpetuates systemic inequality, yet research on bias identification and mitigation remains largely confined to English. Morphologically rich languages such as Greek, where grammatical gender pervades nouns, adjectives, pronouns, and participles, present unique challenges that existing approaches fail to address. This paper elaborates on a systematic methodology primarily focusing on identifying and mitigating gender bias in Greek-language job advertisements and legal documents. To accomplish that task, we define a taxonomy of nine gender bias rules tailored to the linguistic properties of Greek and construct domain-specific annotated datasets comprising 90 expert-curated few-shot examples across both textual domains. Using these resources, we employ XML-structured prompt engineering with in-context learning (ICL)and systematically compare three classes of models: (i) commercial large language models (LLMs), namely Claude Sonnet 4.5 and GPT-5.2, (ii) two open-weight small language models (SLMs), Mistral Small (24B) and Ministral (14B), and (iii) Llama Krikri (8B), a Greek-native language model built on Llama 3.1 and fine-tuned on high-quality Greek corpora. For each input text, the system identifies biased expressions, maps them to specific bias rules, provides explanations, and generates a fully corrected inclusive version. Our experiments reveal substantial performance disparities across model scales and linguistic specialization, with LLMs demonstrating superior contextual reasoning and SLMs exhibiting systematic over-correction and grammatical errors in Greek morphology. We further introduce a critical meta-rule addressing gender agreement with named entities to prevent spurious corrections in legal texts referencing identified individuals. The findings highlight the importance of model scale, language-specific adaptation, and carefully designed prompting strategies for bias mitigation in underrepresented languages. |
Andreas Kontogiannis Ioannis Panageas, Vasilis Pollatos The computational complexity of avoiding strict saddle points in constrained optimization Journal Article arXiv, 2026. @article{Kontogiannis2026b, title = {The computational complexity of avoiding strict saddle points in constrained optimization}, author = {Andreas Kontogiannis, Ioannis Panageas, Vasilis Pollatos}, url = {https://arxiv.org/abs/2604.02285}, doi = {https://doi.org/10.48550/arXiv.2604.02285}, year = {2026}, date = {2026-04-02}, journal = {arXiv}, abstract = {While first-order stationary points (FOSPs) are the traditional targets of non-convex optimization, they often correspond to undesirable strict saddle points. To circumvent this, attention has shifted towards second-order stationary points (SOSPs). In unconstrained settings, finding approximate SOSPs is PLS-complete (Kontogiannis et al.), matching the complexity of finding unconstrained FOSPs (Hollender and Zampetakis). However, the complexity of finding SOSPs in constrained settings remained notoriously unclear and was highlighted as an important open question by both aforementioned works. Under one strict definition, even verifying whether a point is an approximate SOSP is NP-hard (Murty and Kabadi). Under another widely adopted, relaxed definition where non-negative curvature is required only along the null space of the active constraints, the problem lies in TFNP, and algorithms with O(poly(1/epsilon)) running times have been proposed (Lu et al.). In this work, we settle the complexity of constrained SOSP by proving that computing an epsilon-approximate SOSP under the tractable definition is PLS-complete. We demonstrate that our result holds even in the 2D unit square [0,1]^2, and remarkably, even when stationary points are isolated at a distance of Omega(1) from the domain’s boundary. Our result establishes a fundamental barrier: unless PLS is a subset of PPAD (implying PLS = CLS), no deterministic, iterative algorithm with an efficient, continuous update rule can exist for finding approximate SOSPs. This contrasts with the constrained first-order counterpart, for which Fearnley et al. showed that finding an approximate KKT point is CLS-complete. Finally, our result yields the first problem defined in a compact domain to be shown PLS-complete beyond the canonical Real-LocalOpt (Daskalakis and Papadimitriou).”}, keywords = {}, pubstate = {published}, tppubtype = {article} } While first-order stationary points (FOSPs) are the traditional targets of non-convex optimization, they often correspond to undesirable strict saddle points. To circumvent this, attention has shifted towards second-order stationary points (SOSPs). In unconstrained settings, finding approximate SOSPs is PLS-complete (Kontogiannis et al.), matching the complexity of finding unconstrained FOSPs (Hollender and Zampetakis). However, the complexity of finding SOSPs in constrained settings remained notoriously unclear and was highlighted as an important open question by both aforementioned works. Under one strict definition, even verifying whether a point is an approximate SOSP is NP-hard (Murty and Kabadi). Under another widely adopted, relaxed definition where non-negative curvature is required only along the null space of the active constraints, the problem lies in TFNP, and algorithms with O(poly(1/epsilon)) running times have been proposed (Lu et al.). In this work, we settle the complexity of constrained SOSP by proving that computing an epsilon-approximate SOSP under the tractable definition is PLS-complete. We demonstrate that our result holds even in the 2D unit square [0,1]^2, and remarkably, even when stationary points are isolated at a distance of Omega(1) from the domain’s boundary. Our result establishes a fundamental barrier: unless PLS is a subset of PPAD (implying PLS = CLS), no deterministic, iterative algorithm with an efficient, continuous update rule can exist for finding approximate SOSPs. This contrasts with the constrained first-order counterpart, for which Fearnley et al. showed that finding an approximate KKT point is CLS-complete. Finally, our result yields the first problem defined in a compact domain to be shown PLS-complete beyond the canonical Real-LocalOpt (Daskalakis and Papadimitriou)." |
Georgios M Santipantakis Christos Doulkeridis, Petros Brimos Semantic Data Transformation, FAIRification and Provenance for Data Spaces Journal Article Data in Brief, 66 , pp. 112675, 2026, ISSN: 2352-3409. @article{Santipantakis2026, title = {Semantic Data Transformation, FAIRification and Provenance for Data Spaces}, author = {Georgios M Santipantakis, Christos Doulkeridis, Petros Brimos}, url = {https://www.sciencedirect.com/science/article/pii/S2352340926002283}, doi = {https://doi.org/10.1016/j.dib.2026.112675}, issn = {2352-3409}, year = {2026}, date = {2026-03-10}, journal = {Data in Brief}, volume = {66}, pages = {112675}, keywords = {}, pubstate = {published}, tppubtype = {article} } |
Michael Kenteris, Konstantinos Kotis The Convergence of Federated Learning, Knowledge Graphs, and Large Language Models for Language Learning: A Scoping Review Journal Article Applied Sciences, 16 (5), pp. 2611, 2026. @article{Kenteris2026, title = {The Convergence of Federated Learning, Knowledge Graphs, and Large Language Models for Language Learning: A Scoping Review}, author = {Michael Kenteris, Konstantinos Kotis}, url = {https://www.mdpi.com/2076-3417/16/5/2611}, doi = {https://doi.org/10.3390/app16052611}, year = {2026}, date = {2026-03-09}, journal = {Applied Sciences}, volume = {16}, number = {5}, pages = {2611}, abstract = {Large Language Models (LLMs) in Intelligent Computer-Assisted Language Learning enable highly personalized learning, yet raise significant challenges related to pedagogical grounding, data privacy, and instructional validity. Although Knowledge Graphs (KGs) and Federated Learning (FL) can mitigate these issues in isolation, evidence on systematic FL–KG–LLM integration for educational language learning remains limited. This scoping review maps the FL–KG–LLM convergence landscape. Following PRISMA-ScR guidelines, we searched six databases and screened 51 papers (2019–2025) using automated extraction. Our findings indicate limited convergence: no papers integrate all three domains, and 58.8% of approaches remain confined to isolated technological silos. Reporting is also uneven across the corpus, with an average “Not Reported” (NR) rate of 84.5%, most notably for privacy mechanisms (92.2%), validation metrics (90.2%), and Common European Framework of Reference for Languages (CEFR) alignment (88.2%). Domain-specific analysis reveals two distinct patterns: inter-domain gaps (disciplinary silos resulting in expected CEFR absence in single-domain papers) and intra-domain gaps (failure to report domain-critical variables, including 100% parameter NR in FL studies, 86.7% validation NR in KG studies, and 100% CEFR NR in convergence papers). Taken together, these gaps suggest that pedagogical grounding is treated as optional rather than structural. We therefore identify two pillars of pedagogical grounding: a Grounding Pillar, which constrains LLM outputs via Knowledge Graph rules, and a Validation Pillar, which concerns how authoritative frameworks (e.g., CEFR) are mapped onto Knowledge Graph schemas and evaluated. The near-universal absence of CEFR alignment and validation reporting suggests that this second pillar is currently missing, which we term the Integrity Gap—a systematic disconnection between technological innovation and pedagogical grounding inin Intelligent Computer-Assisted Language Learning. By reframing the problem as upstream control and validation, this review informs the design of user-facing automated systems where trust, transparency, and human oversight are critical.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Large Language Models (LLMs) in Intelligent Computer-Assisted Language Learning enable highly personalized learning, yet raise significant challenges related to pedagogical grounding, data privacy, and instructional validity. Although Knowledge Graphs (KGs) and Federated Learning (FL) can mitigate these issues in isolation, evidence on systematic FL–KG–LLM integration for educational language learning remains limited. This scoping review maps the FL–KG–LLM convergence landscape. Following PRISMA-ScR guidelines, we searched six databases and screened 51 papers (2019–2025) using automated extraction. Our findings indicate limited convergence: no papers integrate all three domains, and 58.8% of approaches remain confined to isolated technological silos. Reporting is also uneven across the corpus, with an average “Not Reported” (NR) rate of 84.5%, most notably for privacy mechanisms (92.2%), validation metrics (90.2%), and Common European Framework of Reference for Languages (CEFR) alignment (88.2%). Domain-specific analysis reveals two distinct patterns: inter-domain gaps (disciplinary silos resulting in expected CEFR absence in single-domain papers) and intra-domain gaps (failure to report domain-critical variables, including 100% parameter NR in FL studies, 86.7% validation NR in KG studies, and 100% CEFR NR in convergence papers). Taken together, these gaps suggest that pedagogical grounding is treated as optional rather than structural. We therefore identify two pillars of pedagogical grounding: a Grounding Pillar, which constrains LLM outputs via Knowledge Graph rules, and a Validation Pillar, which concerns how authoritative frameworks (e.g., CEFR) are mapped onto Knowledge Graph schemas and evaluated. The near-universal absence of CEFR alignment and validation reporting suggests that this second pillar is currently missing, which we term the Integrity Gap—a systematic disconnection between technological innovation and pedagogical grounding inin Intelligent Computer-Assisted Language Learning. By reframing the problem as upstream control and validation, this review informs the design of user-facing automated systems where trust, transparency, and human oversight are critical. |
Elias Alevizos Georgios M Santipantakis, Christos Doulkeridis Alexander Artikis Online spatial reasoning for complex event recognition Journal Article GeoInformatica, 30 (1), pp. 9, 2026. @article{Alevizos2026, title = {Online spatial reasoning for complex event recognition}, author = {Elias Alevizos, Georgios M Santipantakis, Christos Doulkeridis, Alexander Artikis}, url = {https://link.springer.com/article/10.1007/s10707-026-00569-z}, doi = {https://doi.org/10.1007/s10707-026-00569-z}, year = {2026}, date = {2026-03-03}, journal = {GeoInformatica}, volume = {30}, number = {1}, pages = {9}, abstract = {Complex Event Recognition (CER) systems have the ability to process streams of events by detecting event patterns with minimal latency. Typically, these patterns have a temporal structure, often resembling the sequential structure of regular expressions. A pattern advances to the next state by checking various conditions on the current and possibly previous events of the stream. CER systems are very efficient in tracking all the possible paths that a pattern may follow and report when a path is complete and a complex event must be reported. In some cases, the conditions that need to be checked may be spatial. For example, in maritime situational awareness, a condition may need to check whether a vessel is close to any other vessel. Such conditions are not easily expressed directly as regular expressions. For such spatio-temporal tasks, there exist dedicated modules which can evaluate this type of conditions efficiently. Thus, we can integrate such a spatio-temporal module within a CER system in order to take advantage of both worlds: the CER engine can accommodate and process complex regular expressions and delegate the evaluation of expensive spatio-temporal tasks to a dedicated module whenever it needs to. We present an approach towards such an integration. We describe how a CER engine, based on symbolic automata, can cooperate with a spatio-temporal link discovery (stLD) module such that the former can leverage the spatio-temporal capabilities of the latter. This cooperation can take place in an online manner rendering the whole system suitable for real-time processing of event streams. We discuss two different communication schemes between the CER engine and the spatio-temporal module and explore when each one should be preferred. We provide a theoretical estimation of the predicted performance of the system under each communication scheme. Our extensive experimental evaluation confirms most of our theoretical predictions.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Complex Event Recognition (CER) systems have the ability to process streams of events by detecting event patterns with minimal latency. Typically, these patterns have a temporal structure, often resembling the sequential structure of regular expressions. A pattern advances to the next state by checking various conditions on the current and possibly previous events of the stream. CER systems are very efficient in tracking all the possible paths that a pattern may follow and report when a path is complete and a complex event must be reported. In some cases, the conditions that need to be checked may be spatial. For example, in maritime situational awareness, a condition may need to check whether a vessel is close to any other vessel. Such conditions are not easily expressed directly as regular expressions. For such spatio-temporal tasks, there exist dedicated modules which can evaluate this type of conditions efficiently. Thus, we can integrate such a spatio-temporal module within a CER system in order to take advantage of both worlds: the CER engine can accommodate and process complex regular expressions and delegate the evaluation of expensive spatio-temporal tasks to a dedicated module whenever it needs to. We present an approach towards such an integration. We describe how a CER engine, based on symbolic automata, can cooperate with a spatio-temporal link discovery (stLD) module such that the former can leverage the spatio-temporal capabilities of the latter. This cooperation can take place in an online manner rendering the whole system suitable for real-time processing of event streams. We discuss two different communication schemes between the CER engine and the spatio-temporal module and explore when each one should be preferred. We provide a theoretical estimation of the predicted performance of the system under each communication scheme. Our extensive experimental evaluation confirms most of our theoretical predictions. |
George Papadopoulos, George Vouros A Learning to maintain safety through expert demonstrations in settings with unknown constraints: A Q-learning perspective Journal Article arXiv, 2026. @article{Papadopoulos2026, title = {Learning to maintain safety through expert demonstrations in settings with unknown constraints: A Q-learning perspective}, author = {George Papadopoulos, George A Vouros}, url = {https://arxiv.org/abs/2602.23816 https://arxiv.org/pdf/2602.23816}, doi = {https://doi.org/10.48550/arXiv.2602.23816}, year = {2026}, date = {2026-02-27}, journal = {arXiv}, abstract = {Given a set of trajectories demonstrating the execution of a task safely in a constrained MDP with observable rewards but with unknown constraints and non-observable costs, we aim to find a policy that maximizes the likelihood of demonstrated trajectories trading the balance between being conservative and increasing significantly the likelihood of high-rewarding trajectories but with potentially unsafe steps. Having these objectives, we aim towards learning a policy that maximizes the probability of the most $promising$ trajectories with respect to the demonstrations. In so doing, we formulate the “promise” of individual state-action pairs in terms of $Q$ values, which depend on task-specific rewards as well as on the assessment of states’ safety, mixing expectations in terms of rewards and safety. This entails a safe Q-learning perspective of the inverse learning problem under constraints: The devised Safe $Q$ Inverse Constrained Reinforcement Learning (SafeQIL) algorithm is compared to state-of-the art inverse constraint reinforcement learning algorithms to a set of challenging benchmark tasks, showing its merits.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Given a set of trajectories demonstrating the execution of a task safely in a constrained MDP with observable rewards but with unknown constraints and non-observable costs, we aim to find a policy that maximizes the likelihood of demonstrated trajectories trading the balance between being conservative and increasing significantly the likelihood of high-rewarding trajectories but with potentially unsafe steps. Having these objectives, we aim towards learning a policy that maximizes the probability of the most $promising$ trajectories with respect to the demonstrations. In so doing, we formulate the “promise" of individual state-action pairs in terms of $Q$ values, which depend on task-specific rewards as well as on the assessment of states’ safety, mixing expectations in terms of rewards and safety. This entails a safe Q-learning perspective of the inverse learning problem under constraints: The devised Safe $Q$ Inverse Constrained Reinforcement Learning (SafeQIL) algorithm is compared to state-of-the art inverse constraint reinforcement learning algorithms to a set of challenging benchmark tasks, showing its merits. |
Dimitrios Doumanas, Konstantinos Kotis ReaDS-KG: An LLM-Knowledge Graph Framework for Reasoned Decision Support in Dynamic Safety-Critical Domains Journal Article TechRxiv, 2026. @article{Doumanas2026, title = {ReaDS-KG: An LLM-Knowledge Graph Framework for Reasoned Decision Support in Dynamic Safety-Critical Domains}, author = {Dimitrios Doumanas, Konstantinos Kotis}, url = {https://www.techrxiv.org/doi/full/10.36227/techrxiv.176826793.34811491/v1}, doi = {https://doi.org/10.36227/techrxiv.176826793.34811491/v1}, year = {2026}, date = {2026-01-13}, journal = {TechRxiv}, abstract = {Safety-critical domains such as military operations, border security, and search-and-rescue must operate under uncertainty, severe time pressure, and continuously changing conditions. In these settings, decision-support systems must not only provide accurate recommendations but also make the underlying reasoning explicit and auditable. This paper introduces ReaDS-KG (Reasoned Decision Support over Knowledge Graphs), an LLM-Knowledge Graph framework that delivers reasoned rather than purely predictive support. ReaDS-KG represents domain knowledge, assets, constraints, and causal dependencies in an ontology-driven knowledge graph, and uses a large language model to (i) translate natural-language questions into Cypher queries, (ii) orchestrate graph-based reasoning over causal structures, and (iii) return narrative answers with explicit justifications grounded in the graph. The framework follows a five-stage pipeline: ontology design, data-to-KG transformation, causal enrichment, LLM-mediated querying, and scenariobased evaluation. To demonstrate its applicability, we instantiate ReaDS-KG in a synthetic brigade-level operational scenario and pose twenty decision-oriented questions, covering feasibility, mobility, sustainment, command-and-control robustness, and risk. We then compare an LLM+KG agent powered by ReaDS-KG to ten active-duty officers using an eight-dimensional scoring rubric. The agent achieves decision-support quality comparable to field-grade officers and clearly above junior officers, while responding at machine response speed and providing transparent reasoning chains. These results suggest that ReaDS-KG can function as a quasi-expert, explainable staff assistant in dynamic safety-critical domains, and the architecture is readily transferable to other safety-critical settings that share similar uncertainty and causal-reasoning requirements, such as border management and disaster response.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Safety-critical domains such as military operations, border security, and search-and-rescue must operate under uncertainty, severe time pressure, and continuously changing conditions. In these settings, decision-support systems must not only provide accurate recommendations but also make the underlying reasoning explicit and auditable. This paper introduces ReaDS-KG (Reasoned Decision Support over Knowledge Graphs), an LLM-Knowledge Graph framework that delivers reasoned rather than purely predictive support. ReaDS-KG represents domain knowledge, assets, constraints, and causal dependencies in an ontology-driven knowledge graph, and uses a large language model to (i) translate natural-language questions into Cypher queries, (ii) orchestrate graph-based reasoning over causal structures, and (iii) return narrative answers with explicit justifications grounded in the graph. The framework follows a five-stage pipeline: ontology design, data-to-KG transformation, causal enrichment, LLM-mediated querying, and scenariobased evaluation. To demonstrate its applicability, we instantiate ReaDS-KG in a synthetic brigade-level operational scenario and pose twenty decision-oriented questions, covering feasibility, mobility, sustainment, command-and-control robustness, and risk. We then compare an LLM+KG agent powered by ReaDS-KG to ten active-duty officers using an eight-dimensional scoring rubric. The agent achieves decision-support quality comparable to field-grade officers and clearly above junior officers, while responding at machine response speed and providing transparent reasoning chains. These results suggest that ReaDS-KG can function as a quasi-expert, explainable staff assistant in dynamic safety-critical domains, and the architecture is readily transferable to other safety-critical settings that share similar uncertainty and causal-reasoning requirements, such as border management and disaster response. |
2025 |
Andreas Soularidis Dimitrios Doumanas, Konstantinos Kotis George Vouros A Automating agentic collaborative ontology engineering with role-playing simulation of LLM-powered agents and RAG technology Journal Article The Knowledge Engineering Review, 40 , pp. e10, 2025. @article{Soularidis2025, title = {Automating agentic collaborative ontology engineering with role-playing simulation of LLM-powered agents and RAG technology}, author = {Andreas Soularidis, Dimitrios Doumanas, Konstantinos Kotis, George A Vouros}, doi = {https://doi.org/10.1017/S026988892510009X}, year = {2025}, date = {2025-12-19}, journal = {The Knowledge Engineering Review}, volume = {40}, pages = {e10}, abstract = {Motivated by the astonishing capabilities of large language models (LLMs) in text-generation, reasoning, and simulation of complex human behaviors, in this paper, we propose a novel multi-component LLM-based framework, namely LLM4ACOE, that fully automates the collaborative ontology engineering (COE) process using role-playing simulation of LLM agents and retrieval augmented generation (RAG) technology. The proposed solution enhances the LLM-powered role-playing simulation with RAG ‘feeding’ the LLM with three different types of external knowledge. This knowledge corresponds to the knowledge required by each of the COE roles (agents), using a component-based framework, as follows: (a) domain-specific data-centric documents, (b) OWL documentation, and (c) ReAct guidelines. The aforementioned components are evaluated in combination, with the aim of investigating their impact on the quality of generated ontologies. The aim of this work is twofold, (a) to identify the capacity of LLM-based agents to generate acceptable (by human-experts) ontologies through agentic collaborative ontology engineering (ACOE) role-playing simulation, at specific levels of acceptance (accuracy, validity, and expressiveness of ontologies) without human intervention and (b) to investigate whether and/or to what extent the selected RAG components affect the quality of the generated ontologies. The evaluation of this novel approach is performed using ChatGPT-o in the domain of search and rescue (SAR) missions. To assess the generated ontologies, quantitative and qualitative measures are employed, focusing on coverage, expressiveness, structure, and human involvement.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Motivated by the astonishing capabilities of large language models (LLMs) in text-generation, reasoning, and simulation of complex human behaviors, in this paper, we propose a novel multi-component LLM-based framework, namely LLM4ACOE, that fully automates the collaborative ontology engineering (COE) process using role-playing simulation of LLM agents and retrieval augmented generation (RAG) technology. The proposed solution enhances the LLM-powered role-playing simulation with RAG ‘feeding’ the LLM with three different types of external knowledge. This knowledge corresponds to the knowledge required by each of the COE roles (agents), using a component-based framework, as follows: (a) domain-specific data-centric documents, (b) OWL documentation, and (c) ReAct guidelines. The aforementioned components are evaluated in combination, with the aim of investigating their impact on the quality of generated ontologies. The aim of this work is twofold, (a) to identify the capacity of LLM-based agents to generate acceptable (by human-experts) ontologies through agentic collaborative ontology engineering (ACOE) role-playing simulation, at specific levels of acceptance (accuracy, validity, and expressiveness of ontologies) without human intervention and (b) to investigate whether and/or to what extent the selected RAG components affect the quality of the generated ontologies. The evaluation of this novel approach is performed using ChatGPT-o in the domain of search and rescue (SAR) missions. To assess the generated ontologies, quantitative and qualitative measures are employed, focusing on coverage, expressiveness, structure, and human involvement. |
Apostolos Glenis, George Vouros Scalable Univariate and Multivariate Time-Series Classifiers with Deep Learning Methods Exploiting Symbolic Representations Journal Article Computers, 14 (12), pp. 563, 2025. @article{Glenis2025, title = {Scalable Univariate and Multivariate Time-Series Classifiers with Deep Learning Methods Exploiting Symbolic Representations}, author = {Apostolos Glenis, George Vouros}, url = {https://www.mdpi.com/2073-431X/14/12/563}, doi = {https://doi.org/10.3390/computers14120563}, year = {2025}, date = {2025-12-17}, journal = {Computers}, volume = {14}, number = {12}, pages = {563}, abstract = {Time-series classification (TSC) is an important task across sciences. Symbolic representations (especially SFA) are very effective at combating noise. In this paper, we employ symbolic representations to create state-of-the-art time-series classifiers, with the aim to advance scalability without sacrificing accuracy. First, we create a graph representation of the time series based on SFA words. We use this representation together with graph kernels and an SVM classifier to create a scalable time-series classifier. Next, we use the graph representation together with a Graph Convolutional Neural Network to test how it fares against state-of-the-art time-series classifiers. Additionally, we devised deep neural networks exploiting the SFA representation, inspired by the text classification domain, to study how they fare against state-of-the-art classifiers. The proposed deep learning classifiers have been adapted and evaluated for the multivariate time-series case and also against state-of-the-art time-series classification algorithms based on symbolic representations.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Time-series classification (TSC) is an important task across sciences. Symbolic representations (especially SFA) are very effective at combating noise. In this paper, we employ symbolic representations to create state-of-the-art time-series classifiers, with the aim to advance scalability without sacrificing accuracy. First, we create a graph representation of the time series based on SFA words. We use this representation together with graph kernels and an SVM classifier to create a scalable time-series classifier. Next, we use the graph representation together with a Graph Convolutional Neural Network to test how it fares against state-of-the-art time-series classifiers. Additionally, we devised deep neural networks exploiting the SFA representation, inspired by the text classification domain, to study how they fare against state-of-the-art classifiers. The proposed deep learning classifiers have been adapted and evaluated for the multivariate time-series case and also against state-of-the-art time-series classification algorithms based on symbolic representations. |
Georgios Bouchouras Dimitrios Doumanas, Andreas Soularidis Konstantinos Kotis George Vouros A Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting Journal Article arXiv, 2025. @article{Bouchouras2025, title = {Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting}, author = {Georgios Bouchouras, Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis, George A Vouros}, url = {https://arxiv.org/pdf/2512.14288}, doi = {https://doi.org/10.48550/arXiv.2512.14288}, year = {2025}, date = {2025-12-16}, journal = {arXiv}, abstract = {This paper explores the integration of Large Language Models (LLMs) in the engineering of a Parkinson’s Disease (PD) monitoring and alerting ontology through four key methodologies: One Shot (OS) prompt techniques, Chain of Thought (CoT) prompts, X-HCOME, and SimX-HCOME+. The primary objective is to determine whether LLMs alone can create comprehensive ontologies and, if not, whether human-LLM collaboration can achieve this goal. Consequently, the paper assesses the effectiveness of LLMs in automated ontology development and the enhancement achieved through human-LLM collaboration. Initial ontology generation was performed using One Shot (OS) and Chain of Thought (CoT) prompts, demonstrating the capability of LLMs to autonomously construct ontologies for PD monitoring and alerting. However, these outputs were not comprehensive and required substantial human refinement to enhance their completeness and accuracy. X-HCOME, a hybrid ontology engineering approach that combines human expertise with LLM capabilities, showed significant improvements in ontology comprehensiveness. This methodology resulted in ontologies that are very similar to those constructed by experts. Further experimentation with SimX-HCOME+, another hybrid methodology emphasizing continuous human supervision and iterative refinement, highlighted the importance of ongoing human involvement. This approach led to the creation of more comprehensive and accurate ontologies. Overall, the paper underscores the potential of human-LLM collaboration in advancing ontology engineering, particularly in complex domains like PD. The results suggest promising directions for future research, including the development of specialized GPT models for ontology construction.}, keywords = {}, pubstate = {published}, tppubtype = {article} } This paper explores the integration of Large Language Models (LLMs) in the engineering of a Parkinson’s Disease (PD) monitoring and alerting ontology through four key methodologies: One Shot (OS) prompt techniques, Chain of Thought (CoT) prompts, X-HCOME, and SimX-HCOME+. The primary objective is to determine whether LLMs alone can create comprehensive ontologies and, if not, whether human-LLM collaboration can achieve this goal. Consequently, the paper assesses the effectiveness of LLMs in automated ontology development and the enhancement achieved through human-LLM collaboration. Initial ontology generation was performed using One Shot (OS) and Chain of Thought (CoT) prompts, demonstrating the capability of LLMs to autonomously construct ontologies for PD monitoring and alerting. However, these outputs were not comprehensive and required substantial human refinement to enhance their completeness and accuracy. X-HCOME, a hybrid ontology engineering approach that combines human expertise with LLM capabilities, showed significant improvements in ontology comprehensiveness. This methodology resulted in ontologies that are very similar to those constructed by experts. Further experimentation with SimX-HCOME+, another hybrid methodology emphasizing continuous human supervision and iterative refinement, highlighted the importance of ongoing human involvement. This approach led to the creation of more comprehensive and accurate ontologies. Overall, the paper underscores the potential of human-LLM collaboration in advancing ontology engineering, particularly in complex domains like PD. The results suggest promising directions for future research, including the development of specialized GPT models for ontology construction. |
Asimina Dimara Konstantinos Kotis, Stamatis Chatzistamatis Nikolaos Evangeliou Chrysaphis Nathanailidis George Tsekouras E Computing, Communications and IoT Applications (ComComAp), 2025, ISBN: 979-8-3315-9143-4. @conference{Dimara2025b, title = {Towards Effective Data Process Pipelines for Legal NLP in English and Non-English Languages: A Greek Case Study}, author = {Asimina Dimara, Konstantinos Kotis, Stamatis Chatzistamatis, Nikolaos Evangeliou, Chrysaphis Nathanailidis, George E Tsekouras}, url = {https://ieeexplore.ieee.org/abstract/document/11353184}, doi = {https://doi.org/10.1109/ComComAp68359.2025.11353184}, isbn = {979-8-3315-9143-4}, year = {2025}, date = {2025-12-14}, booktitle = {Computing, Communications and IoT Applications (ComComAp)}, abstract = {Natural Language Processing (NLP) pipelines form the backbone of legal artificial intelligence applications, yet most existing tools are designed for English corpora and perform poorly when transferred to morphologically rich, non-English languages. This paper investigates these limitations through a comparative study of English and Greek legal texts. It is shown that English-centric pipelines exhibit systematic errors in preprocessing (tokenization, lemmatization, stop-word removal) and fail to capture legal semantics in embeddings, resulting in degraded downstream performance. To address these issues, a generalized framework is proposed that introduces language-specific preprocessing, curated legal resources, and multilingual embeddings fine-tuned on legal corpora. A case study demonstrates how adapted tools substantially improve similarity scores and classification accuracy in Greek legal texts, while highlighting persistent challenges such as grammatical gender bias. The findings underscore the need for fairness-aware, language-specific NLP pipelines to support robust and inclusive legal AI across diverse jurisdictions.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Natural Language Processing (NLP) pipelines form the backbone of legal artificial intelligence applications, yet most existing tools are designed for English corpora and perform poorly when transferred to morphologically rich, non-English languages. This paper investigates these limitations through a comparative study of English and Greek legal texts. It is shown that English-centric pipelines exhibit systematic errors in preprocessing (tokenization, lemmatization, stop-word removal) and fail to capture legal semantics in embeddings, resulting in degraded downstream performance. To address these issues, a generalized framework is proposed that introduces language-specific preprocessing, curated legal resources, and multilingual embeddings fine-tuned on legal corpora. A case study demonstrates how adapted tools substantially improve similarity scores and classification accuracy in Greek legal texts, while highlighting persistent challenges such as grammatical gender bias. The findings underscore the need for fairness-aware, language-specific NLP pipelines to support robust and inclusive legal AI across diverse jurisdictions. |
Alexandros Karakikes, Konstantinos Kotis AI-Assisted OSINT/SOCMINT for Safeguarding Borders: A Systematic Review Journal Article Information, 16 (12), pp. 1095, 2025, ISSN: 2078-2489. @article{Karakikes2025, title = {AI-Assisted OSINT/SOCMINT for Safeguarding Borders: A Systematic Review}, author = {Alexandros Karakikes, Konstantinos Kotis}, url = {https://www.mdpi.com/2078-2489/16/12/1095}, doi = {https://doi.org/10.3390/info16121095}, issn = {2078-2489}, year = {2025}, date = {2025-12-10}, journal = {Information}, volume = {16}, number = {12}, pages = {1095}, abstract = {In the highly volatile realm of global security, the necessity for leading-edge and effectual border resilience tactics has never been more imperative. This PRISMA 2020 guided systematic literature review (SLR) examines the intersection of artificial intelligence (AI), open-source intelligence (OSINT), and social media intelligence (SOCMINT) for enhancing border protection. Our systematic investigation across major databases (IEEE Xplore, Scopus, SpringerLink, MDPI, ACM) and grey literature sources yielded 3932 initial records and, after screening and eligibility assessment, 73 studies and reports from acknowledged organizations, contributing to the evidence synthesis. Three research questions (RQ1–RQ3) were addressed concerning the following: (a) the effectiveness and application of AI in OSINT/SOCMINT for border protection, its (b) data, technical, and operational limitations, and its (c) ethical, legal, and societal implications (GELSI). Evidence matrices summarize the findings, while narrative syntheses underline and thematically group the extracted insights. Results indicate that AI techniques—fluctuating from machine learning (ML) and natural language processing (NLP) to computer vision and emerging large language models (LLMs)—produce quantifiable improvements in forecasting irregular migration, detecting human trafficking, and supporting multimodal intelligence fusion. However, limitations include misinformation, data bias, adversarial vulnerabilities, governance deficits, and sandbox-to-production gaps. Ethical and societal concerns highlight risks of surveillance overreach, discrimination, and insufficient oversight, among others. To our knowledge, this is the first SLR at this intersection. We conclude that, AI-assisted OSINT/SOCMINT presents transformative potential for border protection requiring, nonetheless, balanced governance, robust validation, and future research on LLM/agentic AI, human–AI teaming, and oversight mechanisms.}, keywords = {}, pubstate = {published}, tppubtype = {article} } In the highly volatile realm of global security, the necessity for leading-edge and effectual border resilience tactics has never been more imperative. This PRISMA 2020 guided systematic literature review (SLR) examines the intersection of artificial intelligence (AI), open-source intelligence (OSINT), and social media intelligence (SOCMINT) for enhancing border protection. Our systematic investigation across major databases (IEEE Xplore, Scopus, SpringerLink, MDPI, ACM) and grey literature sources yielded 3932 initial records and, after screening and eligibility assessment, 73 studies and reports from acknowledged organizations, contributing to the evidence synthesis. Three research questions (RQ1–RQ3) were addressed concerning the following: (a) the effectiveness and application of AI in OSINT/SOCMINT for border protection, its (b) data, technical, and operational limitations, and its (c) ethical, legal, and societal implications (GELSI). Evidence matrices summarize the findings, while narrative syntheses underline and thematically group the extracted insights. Results indicate that AI techniques—fluctuating from machine learning (ML) and natural language processing (NLP) to computer vision and emerging large language models (LLMs)—produce quantifiable improvements in forecasting irregular migration, detecting human trafficking, and supporting multimodal intelligence fusion. However, limitations include misinformation, data bias, adversarial vulnerabilities, governance deficits, and sandbox-to-production gaps. Ethical and societal concerns highlight risks of surveillance overreach, discrimination, and insufficient oversight, among others. To our knowledge, this is the first SLR at this intersection. We conclude that, AI-assisted OSINT/SOCMINT presents transformative potential for border protection requiring, nonetheless, balanced governance, robust validation, and future research on LLM/agentic AI, human–AI teaming, and oversight mechanisms. |
Dimitris Kostadimas Vlasios Kasapakis, Konstantinos Kotis Exploiting VR, AIoT and Semantics Towards an Adaptive Virtual Museum Conference 20th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP), 2025, ISBN: 979-8-3315-8704-8. @conference{Kostadimas2025b, title = {Exploiting VR, AIoT and Semantics Towards an Adaptive Virtual Museum}, author = {Dimitris Kostadimas, Vlasios Kasapakis, Konstantinos Kotis}, url = {https://ieeexplore.ieee.org/abstract/document/11309793}, doi = {https://doi.org/10.1109/SMAP66932.2025.00034}, isbn = {979-8-3315-8704-8}, year = {2025}, date = {2025-11-27}, booktitle = {20th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP)}, pages = {157-162}, abstract = {Museums have long been spaces of wonder and discovery, but as technology evolves, so do the ways we engage with these cultural treasures. The design of adaptive virtual environments becomes essential to maintaining user interest and relevance. In this paper, an adaptive virtual museum system is proposed that explores the use of virtual reality (VR), artificial intelligence (AI), Internet of Things (IoT) as well as semantics to personalize and optimize virtual exhibition experiences. Based on the results of our previous research conducted regarding the possible combination of VR, AI and IoT (AIoT) for the design of innovative intelligent systems in different domains, our current work proposes a novel way to integrate all these technologies within the domain of cultural heritage (CH), a combination that remains relatively underexplored. The proposed framework, which is currently a work in progress, introduces new ways to modeling museums’ visitor behavior and preferences (mainly by using head-mounted displays (HMDs)) in a VR environment to dynamically adapt exhibition layouts, as well as to provide personalized content through a digital twin (DT) of a real museum. A key focus lies in intelligent user profiling and route/layout optimization to enhance visitor engagement and provide rich content through integration of Large Language Models (LLM). Although implementation is ongoing, this paper describes the conceptual design, core objectives, and anticipated impact on the broader scope of adaptive multimedia applications and personalized cultural experiences.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Museums have long been spaces of wonder and discovery, but as technology evolves, so do the ways we engage with these cultural treasures. The design of adaptive virtual environments becomes essential to maintaining user interest and relevance. In this paper, an adaptive virtual museum system is proposed that explores the use of virtual reality (VR), artificial intelligence (AI), Internet of Things (IoT) as well as semantics to personalize and optimize virtual exhibition experiences. Based on the results of our previous research conducted regarding the possible combination of VR, AI and IoT (AIoT) for the design of innovative intelligent systems in different domains, our current work proposes a novel way to integrate all these technologies within the domain of cultural heritage (CH), a combination that remains relatively underexplored. The proposed framework, which is currently a work in progress, introduces new ways to modeling museums’ visitor behavior and preferences (mainly by using head-mounted displays (HMDs)) in a VR environment to dynamically adapt exhibition layouts, as well as to provide personalized content through a digital twin (DT) of a real museum. A key focus lies in intelligent user profiling and route/layout optimization to enhance visitor engagement and provide rich content through integration of Large Language Models (LLM). Although implementation is ongoing, this paper describes the conceptual design, core objectives, and anticipated impact on the broader scope of adaptive multimedia applications and personalized cultural experiences. |
Andreas Sideras Konstantinos Bougiatiotis, Elias Zavitsanos Georgios Paliouras George Vouros A Multimodal Alignment-Based Anomaly Detection Method for Bankruptcy Prediction Conference Proceedings of the 6th ACM International Conference on AI in Finance, 2025, ISBN: 9798400722202. @conference{Sideras2025, title = {A Multimodal Alignment-Based Anomaly Detection Method for Bankruptcy Prediction}, author = {Andreas Sideras, Konstantinos Bougiatiotis, Elias Zavitsanos, Georgios Paliouras, George Vouros}, url = {https://dl.acm.org/doi/full/10.1145/3768292.3770380}, doi = {https://doi.org/10.1145/3768292.3770380}, isbn = {9798400722202}, year = {2025}, date = {2025-11-15}, booktitle = {Proceedings of the 6th ACM International Conference on AI in Finance}, pages = {53-61}, abstract = {We present a novel anomaly detection method for next-year bankruptcy prediction, utilizing a combination of financial figures and textual content from annual reports. Our approach, MABAD, learns a shared representation space where non-bankrupt firms share position and orientation. Samples that deviate from this pattern are assigned a higher anomaly score. The proposed method is tailored for highly imbalanced scenarios and is robust to heterogeneous, incomplete, and potentially contradictory inputs. We demonstrate that MABAD consistently outperforms a range of strong baselines, and we also curate and release a new publicly available multisource dataset to foster further research in the domain.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } We present a novel anomaly detection method for next-year bankruptcy prediction, utilizing a combination of financial figures and textual content from annual reports. Our approach, MABAD, learns a shared representation space where non-bankrupt firms share position and orientation. Samples that deviate from this pattern are assigned a higher anomaly score. The proposed method is tailored for highly imbalanced scenarios and is robust to heterogeneous, incomplete, and potentially contradictory inputs. We demonstrate that MABAD consistently outperforms a range of strong baselines, and we also curate and release a new publicly available multisource dataset to foster further research in the domain. |
Elias Zavitsanos Konstantinos Bougiatiotis, Andreas Sideras Georgios Paliouras Positive-Unlabeled Learning for Financial Misstatement Detection under Realistic Constraints Conference ICAIF ’25: Proceedings of the 6th ACM International Conference on AI in Finance, 2025, ISBN: 9798400722202. @conference{Zavitsanos2025, title = {Positive-Unlabeled Learning for Financial Misstatement Detection under Realistic Constraints}, author = {Elias Zavitsanos, Konstantinos Bougiatiotis, Andreas Sideras, Georgios Paliouras}, url = {https://dl.acm.org/doi/full/10.1145/3768292.3770366 https://dl.acm.org/doi/epdf/10.1145/3768292.3770366}, doi = {https://doi.org/10.1145/3768292.3770366}, isbn = {9798400722202}, year = {2025}, date = {2025-11-15}, booktitle = {ICAIF ’25: Proceedings of the 6th ACM International Conference on AI in Finance}, pages = {864-872}, abstract = {Detecting financial misstatements is critical for market integrity but remains challenging due to class imbalance, delayed discovery, and limited labeled data. We propose a novel Positive-Unlabeled (PU) learning framework that models the detection task under realistic constraints, where only a small subset of misstatements is known at training time. Our approach integrates unlabeled data into training, preserves temporal structure, and accounts for extreme imbalance. We construct and release a benchmark dataset reflecting these characteristics and evaluate several PU learning methods against recent baselines. Results show that PU-based models consistently outperform supervised approaches, highlighting their suitability for real-world misstatement detection.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Detecting financial misstatements is critical for market integrity but remains challenging due to class imbalance, delayed discovery, and limited labeled data. We propose a novel Positive-Unlabeled (PU) learning framework that models the detection task under realistic constraints, where only a small subset of misstatements is known at training time. Our approach integrates unlabeled data into training, preserves temporal structure, and accounts for extreme imbalance. We construct and release a benchmark dataset reflecting these characteristics and evaluate several PU learning methods against recent baselines. Results show that PU-based models consistently outperform supervised approaches, highlighting their suitability for real-world misstatement detection. |
Dimitrios Doumanas Andreas Soularidis, Konstantinos Kotis Causal Reasoning and Large Language Models for Military Decision-Making: Rethinking the Command Structures in the Era of Generative AI Journal Article AI, 7 (1), pp. 14, 2025, ISSN: 2673-2688. @article{Doumanas2025e, title = {Causal Reasoning and Large Language Models for Military Decision-Making: Rethinking the Command Structures in the Era of Generative AI}, author = {Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis}, url = {https://www.mdpi.com/2673-2688/7/1/14}, doi = {https://doi.org/10.3390/ai7010014}, issn = {2673-2688}, year = {2025}, date = {2025-10-24}, journal = {AI}, volume = {7}, number = {1}, pages = {14}, abstract = {Military decision-making is inherently complex and highly critical, requiring commanders to assess multiple variables in real-time, anticipate second-order effects, and adapt strategies based on continuously evolving battlefield conditions. Traditional approaches rely on domain expertise, experience, and intuition, often supported by decision-support systems designed by military experts. With the rapid advancement of Large Language Models (LLMs) such as ChatGPT, Claude, and DeepSeek, a new research question emerges: can LLMs perform causal reasoning at a level that could meaningfully replace human decision-makers, or should they remain human-led decision-support tools in high-stakes environments? This paper explores the causal reasoning capabilities of LLMs for operational and strategic military decisions. Unlike conventional AI models that rely primarily on correlation-based predictions, LLMs are now able to engage in multi-perspective reasoning, intervention analysis, and scenario-based assessments. We introduce a structured empirical evaluation framework to assess LLM performance through 10 de-identified real-world-inspired battle scenarios, ensuring models reason over provided inputs rather than memorized data. Critically, LLM outputs are systematically compared against a human expert baseline, composed of military officers across multiple ranks and years of operational experience. The evaluation focuses on precision, recall, causal reasoning depth, adaptability, and decision soundness. Our findings provide a rigorous comparative assessment of whether carefully prompted LLMs can assist, complement, or approach expert-level performance in military planning. While fully autonomous AI-led command remains premature, the results suggest that LLMs can offer valuable support in complex decision processes when integrated as part of hybrid human-AI decision-support frameworks. Since our evaluation directly tests this capability, this paradigm shift raises fundamental question: Is there a possibility to fully replace high-ranking officers/commanders in leading critical military operations, or should AI-driven tools remain as decision-support systems enhancing human-driven battlefield strategies?}, keywords = {}, pubstate = {published}, tppubtype = {article} } Military decision-making is inherently complex and highly critical, requiring commanders to assess multiple variables in real-time, anticipate second-order effects, and adapt strategies based on continuously evolving battlefield conditions. Traditional approaches rely on domain expertise, experience, and intuition, often supported by decision-support systems designed by military experts. With the rapid advancement of Large Language Models (LLMs) such as ChatGPT, Claude, and DeepSeek, a new research question emerges: can LLMs perform causal reasoning at a level that could meaningfully replace human decision-makers, or should they remain human-led decision-support tools in high-stakes environments? This paper explores the causal reasoning capabilities of LLMs for operational and strategic military decisions. Unlike conventional AI models that rely primarily on correlation-based predictions, LLMs are now able to engage in multi-perspective reasoning, intervention analysis, and scenario-based assessments. We introduce a structured empirical evaluation framework to assess LLM performance through 10 de-identified real-world-inspired battle scenarios, ensuring models reason over provided inputs rather than memorized data. Critically, LLM outputs are systematically compared against a human expert baseline, composed of military officers across multiple ranks and years of operational experience. The evaluation focuses on precision, recall, causal reasoning depth, adaptability, and decision soundness. Our findings provide a rigorous comparative assessment of whether carefully prompted LLMs can assist, complement, or approach expert-level performance in military planning. While fully autonomous AI-led command remains premature, the results suggest that LLMs can offer valuable support in complex decision processes when integrated as part of hybrid human-AI decision-support frameworks. Since our evaluation directly tests this capability, this paradigm shift raises fundamental question: Is there a possibility to fully replace high-ranking officers/commanders in leading critical military operations, or should AI-driven tools remain as decision-support systems enhancing human-driven battlefield strategies? |
Theodore Tranos Nikolaos Fesakis, Thomas Vasileiou Sotirios Christopoulos Georgio Loukos Maria Koutsoupidou 2025 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe), IEEE, 2025, ISBN: 979-8-3315-2503-3. @conference{Tranos2025, title = {AI-Based Energy Forecasting at Different Distribution Grid Levels to Support Baseline Definition and DSO Participation in LFMs}, author = {Theodore Tranos, Nikolaos Fesakis, Thomas Vasileiou, Sotirios Christopoulos, Georgio Loukos, Maria Koutsoupidou}, url = {https://ieeexplore.ieee.org/abstract/document/11305676}, doi = {https://doi.org/10.1109/ISGTEurope64741.2025.11305676}, isbn = {979-8-3315-2503-3}, year = {2025}, date = {2025-10-20}, booktitle = {2025 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe)}, pages = {1-5}, publisher = {IEEE}, abstract = {A crucial aspect of Local Flexibility Markets (LFMs) is the definition of a baseline for energy production and demand forecasting, which serves as a reference for validating and compensating flexibility services. In this study, we explore the application of machine learning techniques, specifically Long Short-Term Memory (LSTM) networks, to establish accurate baselines for consumers and producers connected to the LV grid. The LSTM models leverage real historical demand and generation data from DSO smart meters in Mesogeia, Greece, combined with weather variables such as temperature and cloud coverage, to enhance forecasting accuracy. Our goal is to evaluate forecasting accuracy at the individual participant level and compare it with the accuracy obtained from forecasting on aggregated consumption/production data within a specific grid segment or using data from the secondary substation to which the participants are connected.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } A crucial aspect of Local Flexibility Markets (LFMs) is the definition of a baseline for energy production and demand forecasting, which serves as a reference for validating and compensating flexibility services. In this study, we explore the application of machine learning techniques, specifically Long Short-Term Memory (LSTM) networks, to establish accurate baselines for consumers and producers connected to the LV grid. The LSTM models leverage real historical demand and generation data from DSO smart meters in Mesogeia, Greece, combined with weather variables such as temperature and cloud coverage, to enhance forecasting accuracy. Our goal is to evaluate forecasting accuracy at the individual participant level and compare it with the accuracy obtained from forecasting on aggregated consumption/production data within a specific grid segment or using data from the secondary substation to which the participants are connected. |
Asimina Dimara Konstantinos Kotis, Alexios Papaioannou Stamatis Chatzistamatis Nikolaos Evangeliou Chrysaphis Nathanailidis George Tsekouras 5th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), 2025, ISBN: 979-8-3315-3556-8. @conference{Dimara2025, title = {Data Collection, Organization, and Privacy-Preserving Preparation for Edge-Based LLMs in Legal Text Analytics}, author = {Asimina Dimara, Konstantinos Kotis, Alexios Papaioannou, Stamatis Chatzistamatis, Nikolaos Evangeliou, Chrysaphis Nathanailidis, George Tsekouras}, url = {https://ieeexplore.ieee.org/abstract/document/11277858}, doi = {https://doi.org/10.1109/ICECCME64568.2025.11277858}, isbn = {979-8-3315-3556-8}, year = {2025}, date = {2025-10-16}, booktitle = {5th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME)}, abstract = {Providing fairness and privacy in automated legal text processing is an essential issue, especially with the increasing usage of Large Language Models (LLMs), in sensitive public sector applications. This paper presents a modular edge native domain-specific architecture for legal document processing that avoids cloud infrastructure and external APIs. The system combines local ingestion, semantic embedding, and retrievalaugmented generation to empower autonomous agents for applications such as bias detection and clause summarization. Inference is done exclusively on-device by a 4-bit quantized LLaMA model run by CPU-only runtimes. Tested on the CLEAR-Bias benchmark, the system gets 92% prompt relevance and 90% output coherence, inference latency below 6.5 s, and memory usage below 5.5 GB. These findings validate the effectiveness of privacy-preserving, regulation-conforming legal NLP in constrained environments.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Providing fairness and privacy in automated legal text processing is an essential issue, especially with the increasing usage of Large Language Models (LLMs), in sensitive public sector applications. This paper presents a modular edge native domain-specific architecture for legal document processing that avoids cloud infrastructure and external APIs. The system combines local ingestion, semantic embedding, and retrievalaugmented generation to empower autonomous agents for applications such as bias detection and clause summarization. Inference is done exclusively on-device by a 4-bit quantized LLaMA model run by CPU-only runtimes. Tested on the CLEAR-Bias benchmark, the system gets 92% prompt relevance and 90% output coherence, inference latency below 6.5 s, and memory usage below 5.5 GB. These findings validate the effectiveness of privacy-preserving, regulation-conforming legal NLP in constrained environments. |
Eleftherios Efkleidis Stefanou Pavlos Bitilis, Georgios Bouchouras Konstantinos Kotis Collecting, Integrating and Processing IoT Sensor Data on Edge Devices for PD Monitoring: A Scoping Review Journal Article Applied Sciences, 15 (19), pp. 10541, 2025, ISSN: 2076-3417. @article{Stefanou2025b, title = {Collecting, Integrating and Processing IoT Sensor Data on Edge Devices for PD Monitoring: A Scoping Review}, author = {Eleftherios Efkleidis Stefanou, Pavlos Bitilis, Georgios Bouchouras, Konstantinos Kotis}, url = {https://www.mdpi.com/2076-3417/15/19/10541}, doi = {https://doi.org/10.3390/app151910541}, issn = {2076-3417}, year = {2025}, date = {2025-09-29}, journal = {Applied Sciences}, volume = {15}, number = {19}, pages = {10541}, abstract = {Bradykinesia and tremor are critical motor symptoms in diagnosing and monitoring Parkinson’s disease (PD), a progressive neurodegenerative disorder. The integration of IoT sensors, smartwatch technology, and edge computing has facilitated real-time collection, processing, and analysis of data related to these impairments, enabling continuous monitoring of PD beyond traditional clinical settings. This survey provides a comprehensive review of recent technological advancements in data collection from wearable IoT sensors and its semantic integration and processing on edge devices, emphasizing methods optimized for efficient and low-latency processing. Additionally, this survey explores AI-driven techniques for detecting and analyzing bradykinesia and tremor symptoms on edge devices. By leveraging localized computation on edge devices, these approaches facilitate energy efficiency, data privacy, and scalability, making them suitable for deployment in real environments. This paper also examines related open-source tools and datasets, assessing their roles in improving reproducibility and integration into these environments. Furthermore, key challenges, including variability in real environments, model generalization, and computational constraints, are discussed, along with potential strategies to enhance detection accuracy and system robustness. By bridging the gap between sensor data collection and integration, and AI-based detection of bradykinesia and tremor on edge devices, this survey intends to contribute to the development of efficient, scalable, and privacy-preserving healthcare solutions for continuous PD monitoring.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Bradykinesia and tremor are critical motor symptoms in diagnosing and monitoring Parkinson’s disease (PD), a progressive neurodegenerative disorder. The integration of IoT sensors, smartwatch technology, and edge computing has facilitated real-time collection, processing, and analysis of data related to these impairments, enabling continuous monitoring of PD beyond traditional clinical settings. This survey provides a comprehensive review of recent technological advancements in data collection from wearable IoT sensors and its semantic integration and processing on edge devices, emphasizing methods optimized for efficient and low-latency processing. Additionally, this survey explores AI-driven techniques for detecting and analyzing bradykinesia and tremor symptoms on edge devices. By leveraging localized computation on edge devices, these approaches facilitate energy efficiency, data privacy, and scalability, making them suitable for deployment in real environments. This paper also examines related open-source tools and datasets, assessing their roles in improving reproducibility and integration into these environments. Furthermore, key challenges, including variability in real environments, model generalization, and computational constraints, are discussed, along with potential strategies to enhance detection accuracy and system robustness. By bridging the gap between sensor data collection and integration, and AI-based detection of bradykinesia and tremor on edge devices, this survey intends to contribute to the development of efficient, scalable, and privacy-preserving healthcare solutions for continuous PD monitoring. |
Andreas Kontogiannis Vasilis Pollatos, Gabriele Farina Panayotis Mertikopoulos Ioannis Panageas The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025 poster), 2025. @conference{Kontogiannis2025b, title = {Efficient kernelized learning in polyhedral games beyond full-information: From Colonel Blotto to congestion games}, author = {Andreas Kontogiannis, Vasilis Pollatos, Gabriele Farina, Panayotis Mertikopoulos, Ioannis Panageas}, url = {https://openreview.net/attachment?id=FUBaZDMOFj&name=pdf https://arxiv.org/pdf/2509.20919}, doi = {https://doi.org/10.48550/arXiv.2509.20919}, year = {2025}, date = {2025-09-25}, booktitle = {The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025 poster)}, journal = {The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025 poster)}, abstract = {We examine the problem of efficiently learning coarse correlated equilibria (CCE) in polyhedral games, that is, normal-form games with an exponentially large number of actions per player and an underlying combinatorial structure—such as the classic Colonel Blotto game or congestion games. Achieving computational efficiency in this setting requires learning algorithms whose regret and per-iteration complexity scale at most polylogarithmically with the size of the players’ action sets. This challenge has recently been addressed in the full-information setting, primarily through the use of kernelization; however, in the more realistic partial information setting, the situation is much more challenging, and existing approaches result in suboptimal and impractical runtime complexity to learn CCE. We address this gap via a novel kernelization-based framework for payoff-based learning in polyhedral games, which we then apply to certain key classes of polyhedral games—namely Colonel Blotto, graphic matroid and network congestion games. In so doing, we obtain a range of computationally efficient payoff-based learning algorithms which significantly improve upon prior work in terms of the runtime for learning CCE.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } We examine the problem of efficiently learning coarse correlated equilibria (CCE) in polyhedral games, that is, normal-form games with an exponentially large number of actions per player and an underlying combinatorial structure—such as the classic Colonel Blotto game or congestion games. Achieving computational efficiency in this setting requires learning algorithms whose regret and per-iteration complexity scale at most polylogarithmically with the size of the players’ action sets. This challenge has recently been addressed in the full-information setting, primarily through the use of kernelization; however, in the more realistic partial information setting, the situation is much more challenging, and existing approaches result in suboptimal and impractical runtime complexity to learn CCE. We address this gap via a novel kernelization-based framework for payoff-based learning in polyhedral games, which we then apply to certain key classes of polyhedral games—namely Colonel Blotto, graphic matroid and network congestion games. In so doing, we obtain a range of computationally efficient payoff-based learning algorithms which significantly improve upon prior work in terms of the runtime for learning CCE. |
Adam Koletis Pavlos Bitilis, Georgios Bouchouras Konstantinos Kotis Information, 16 (9), pp. 820, 2025, ISSN: 2078-2489. @article{Koletis2025, title = {A Comparative Analysis of Parkinson’s Disease Diagnosis Approaches Using Drawing-Based Datasets: Utilizing Large Language Models, Machine Learning, and Fuzzy Ontologies}, author = {Adam Koletis, Pavlos Bitilis, Georgios Bouchouras, Konstantinos Kotis}, url = {https://www.mdpi.com/2078-2489/16/9/820}, doi = {https://doi.org/10.3390/info16090820}, issn = {2078-2489}, year = {2025}, date = {2025-09-22}, journal = {Information}, volume = {16}, number = {9}, pages = {820}, abstract = {Parkinson’s disease (PD) is a progressive neurodegenerative disorder that impairs motor function, often causing tremors and difficulty with movement control. A promising diagnostic method involves analyzing hand-drawn patterns, such as spirals and waves, which show characteristic distortions in individuals with PD. This study compares three computational approaches for classifying individuals as Parkinsonian or healthy based on drawing-derived features: (1) Large Language Models (LLMs), (2) traditional machine learning (ML) algorithms, and (3) a fuzzy ontology-based method using fuzzy sets and Fuzzy-OWL2. Each method offers unique strengths: LLMs leverage pre-trained knowledge for subtle pattern detection, ML algorithms excel in feature extraction and predictive accuracy, and fuzzy ontologies provide interpretable, logic-based reasoning under uncertainty. Using three structured handwriting datasets of varying complexity, we assessed performance in terms of accuracy, interpretability, and generalization. Among the approaches, the fuzzy ontology-based method showed the strongest performance on complex tasks, achieving a high F1-score, while ML models demonstrated strong generalization and LLMs offered a reliable, interpretable baseline. These findings suggest that combining symbolic and statistical AI may improve drawing-based PD diagnosis.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Parkinson’s disease (PD) is a progressive neurodegenerative disorder that impairs motor function, often causing tremors and difficulty with movement control. A promising diagnostic method involves analyzing hand-drawn patterns, such as spirals and waves, which show characteristic distortions in individuals with PD. This study compares three computational approaches for classifying individuals as Parkinsonian or healthy based on drawing-derived features: (1) Large Language Models (LLMs), (2) traditional machine learning (ML) algorithms, and (3) a fuzzy ontology-based method using fuzzy sets and Fuzzy-OWL2. Each method offers unique strengths: LLMs leverage pre-trained knowledge for subtle pattern detection, ML algorithms excel in feature extraction and predictive accuracy, and fuzzy ontologies provide interpretable, logic-based reasoning under uncertainty. Using three structured handwriting datasets of varying complexity, we assessed performance in terms of accuracy, interpretability, and generalization. Among the approaches, the fuzzy ontology-based method showed the strongest performance on complex tasks, achieving a high F1-score, while ML models demonstrated strong generalization and LLMs offered a reliable, interpretable baseline. These findings suggest that combining symbolic and statistical AI may improve drawing-based PD diagnosis. |
Theocharis Kravaris, George Vouros A Transferable aircraft trajectory prediction with generative deep imitation learning Journal Article International Journal of Data Science and Analytics, 20 (3), pp. 1977-1999, 2025. @article{Kravaris2025, title = {Transferable aircraft trajectory prediction with generative deep imitation learning}, author = {Theocharis Kravaris, George A Vouros}, url = {https://link.springer.com/article/10.1007/s41060-024-00574-1}, doi = {https://doi.org/10.1007/s41060-024-00574-1}, year = {2025}, date = {2025-09-01}, journal = {International Journal of Data Science and Analytics}, volume = {20}, number = {3}, pages = {1977-1999}, abstract = {Trajectory-oriented transformations to air traffic management operations require high fidelity aircraft trajectory prediction capabilities. Data-driven trajectory prediction approaches provide promising results, albeit with important limitations that hinder seriously the efficient and effective deployment of trajectory prediction methods: They need abundant training effort with a large amount of training samples and require training distinct models for different origin–destination (OD) airport pairs. In this paper, we address the problem of building transferable trajectory prediction models, casting the prediction problem as a transferable imitation task, introducing a novel formulation which (a) provides the capability to utilize trained models, in new OD pairs, offering a warm starting for computationally efficient training, and (b) improves the efficacy of data-driven trajectory prediction. The proposed approach provides very accurate results for large look-ahead time predictions, even if transferable models have been trained with few samples.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Trajectory-oriented transformations to air traffic management operations require high fidelity aircraft trajectory prediction capabilities. Data-driven trajectory prediction approaches provide promising results, albeit with important limitations that hinder seriously the efficient and effective deployment of trajectory prediction methods: They need abundant training effort with a large amount of training samples and require training distinct models for different origin–destination (OD) airport pairs. In this paper, we address the problem of building transferable trajectory prediction models, casting the prediction problem as a transferable imitation task, introducing a novel formulation which (a) provides the capability to utilize trained models, in new OD pairs, offering a warm starting for computationally efficient training, and (b) improves the efficacy of data-driven trajectory prediction. The proposed approach provides very accurate results for large look-ahead time predictions, even if transferable models have been trained with few samples. |
Dimitrios Doumanas Alexandros Karakikes, Andreas Soularidis Efstathios Mainas Konstantinos Kotis Emerging Threat Vectors: How Malicious Actors Exploit LLMs to Undermine Border Security Journal Article AI, 6 (9), pp. 232, 2025, ISSN: 2673-2688. @article{Doumanas2025d, title = {Emerging Threat Vectors: How Malicious Actors Exploit LLMs to Undermine Border Security}, author = {Dimitrios Doumanas, Alexandros Karakikes, Andreas Soularidis, Efstathios Mainas, Konstantinos Kotis}, url = {https://www.mdpi.com/2673-2688/6/9/232}, doi = {https://doi.org/10.3390/ai6090232}, issn = {2673-2688}, year = {2025}, date = {2025-09-01}, journal = {AI}, volume = {6}, number = {9}, pages = {232}, abstract = {The rapid proliferation of Large Language Models (LLMs) has democratized access to advanced generative capabilities while raising urgent concerns about misuse in sensitive security domains. Border security, in particular, represents a high-risk environment where malicious actors may exploit LLMs for document forgery, synthetic identity creation, logistics planning, or disinformation campaigns. Existing studies often highlight such risks in theory, yet few provide systematic empirical evidence of how state-of-the-art LLMs can be exploited. This paper introduces the Silent Adversary Framework (SAF), a structured pipeline that models the sequential stages by which obfuscated prompts can covertly bypass safeguards. We evaluate ten high-risk scenarios using five leading models—GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Flash, Grok 3, and Runway Gen-2—and assess outputs through three standardized metrics: Bypass Success Rate (BSR), Output Realism Score (ORS), and Operational Risk Level (ORL). Results reveal that, while all models exhibited some susceptibility, vulnerabilities were heterogeneous. Claude showed greater resistance in chemistry-related prompts, whereas GPT-4o and Gemini generated highly realistic outputs in identity fraud and logistics optimization tasks. Document forgery attempts produced only partially successful templates that lacked critical security features. These findings highlight the uneven distribution of risks across models and domains. By combining a reproducible adversarial framework with empirical testing, this study advances the evidence base on LLM misuse and provides actionable insights for policymakers and border security agencies, underscoring the need for stronger safeguards and oversight in the deployment of generative AI.}, keywords = {}, pubstate = {published}, tppubtype = {article} } The rapid proliferation of Large Language Models (LLMs) has democratized access to advanced generative capabilities while raising urgent concerns about misuse in sensitive security domains. Border security, in particular, represents a high-risk environment where malicious actors may exploit LLMs for document forgery, synthetic identity creation, logistics planning, or disinformation campaigns. Existing studies often highlight such risks in theory, yet few provide systematic empirical evidence of how state-of-the-art LLMs can be exploited. This paper introduces the Silent Adversary Framework (SAF), a structured pipeline that models the sequential stages by which obfuscated prompts can covertly bypass safeguards. We evaluate ten high-risk scenarios using five leading models—GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Flash, Grok 3, and Runway Gen-2—and assess outputs through three standardized metrics: Bypass Success Rate (BSR), Output Realism Score (ORS), and Operational Risk Level (ORL). Results reveal that, while all models exhibited some susceptibility, vulnerabilities were heterogeneous. Claude showed greater resistance in chemistry-related prompts, whereas GPT-4o and Gemini generated highly realistic outputs in identity fraud and logistics optimization tasks. Document forgery attempts produced only partially successful templates that lacked critical security features. These findings highlight the uneven distribution of risks across models and domains. By combining a reproducible adversarial framework with empirical testing, this study advances the evidence base on LLM misuse and provides actionable insights for policymakers and border security agencies, underscoring the need for stronger safeguards and oversight in the deployment of generative AI. |
Konstantinos Kotis Eleni Angoura, Eleni-Ioanna Lyngri Emerging technologies in smart libraries for visually impaired people: challenges and design considerations Journal Article ACM Journal on Computing and Cultural Heritage, 18 (3), pp. 1-37, 2025, ISSN: 1556-4673. @article{Kotis2025, title = {Emerging technologies in smart libraries for visually impaired people: challenges and design considerations}, author = {Konstantinos Kotis, Eleni Angoura, Eleni-Ioanna Lyngri}, url = {https://dl.acm.org/doi/full/10.1145/3727965}, doi = {https://doi.org/10.1145/3727965}, issn = {1556-4673}, year = {2025}, date = {2025-07-24}, journal = {ACM Journal on Computing and Cultural Heritage}, volume = {18}, number = {3}, pages = {1-37}, abstract = {Emerging technologies are transforming cultural spaces in a variety of ways, presenting opportunities and challenges. Autonomous robots, eXtended Reality, AI, Digital Twins, and Internet of Things are only a few examples of such technologies, with accessibility and inclusivity of people to these technologies to be considered key challenges. In general, the use of emerging technologies in cultural spaces presents exciting opportunities for enhancing visitors’ experience and engaging new participants. However, it is important to also consider the inclusion ability of people with special needs and to ensure that these emerging technologies are used in an accessible-to-all and inclusive way. The aim of this article is to review the state-of-the-art and current trends in approaches that use emerging technologies in the domain of smart libraries designed to include visually impaired people in a common innovative way for the whole community of visitors, discuss open issues and challenges identified in such a cultural environment/case, and propose a novel approach based on specific design considerations of the specific domain.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Emerging technologies are transforming cultural spaces in a variety of ways, presenting opportunities and challenges. Autonomous robots, eXtended Reality, AI, Digital Twins, and Internet of Things are only a few examples of such technologies, with accessibility and inclusivity of people to these technologies to be considered key challenges. In general, the use of emerging technologies in cultural spaces presents exciting opportunities for enhancing visitors’ experience and engaging new participants. However, it is important to also consider the inclusion ability of people with special needs and to ensure that these emerging technologies are used in an accessible-to-all and inclusive way. The aim of this article is to review the state-of-the-art and current trends in approaches that use emerging technologies in the domain of smart libraries designed to include visually impaired people in a common innovative way for the whole community of visitors, discuss open issues and challenges identified in such a cultural environment/case, and propose a novel approach based on specific design considerations of the specific domain. |
Sotiris Angelis Joana Pinho, Athanasia Sykiotou Dimitar Markov Stamatis Chatzistamatis Stamatis Spirou George Tsekouras Konstantinos Kotis RRAO: An Ontology for the Representation of Reoffending Risk Assessment Knowledge Conference 16th International Conference on Information, Intelligence, Systems & Applications (IISA), 2025, ISBN: 979-8-3315-5636-5. @conference{Angelis2025, title = {RRAO: An Ontology for the Representation of Reoffending Risk Assessment Knowledge}, author = {Sotiris Angelis, Joana Pinho, Athanasia Sykiotou, Dimitar Markov, Stamatis Chatzistamatis, Stamatis Spirou, George Tsekouras, Konstantinos Kotis}, doi = {https://doi.org/10.1109/IISA66859.2025.11311249}, isbn = {979-8-3315-5636-5}, year = {2025}, date = {2025-07-10}, booktitle = {16th International Conference on Information, Intelligence, Systems & Applications (IISA)}, pages = {1-9}, abstract = {Judicial decision making related to parole, sentencing, rehabilitation, reintegration, and public safety is often supported by the assessment of the risk of reoffending. AI prediction systems can introduce bias in the analysis of reoffending riskrelated data. Several studies criticize the fairness of such AI systems. This paper presents the Reoffending Risk Assessment Ontology (RRAO) which aims to provide a comprehensive representation of reoffending risk and recidivism knowledge integrated into ontology-based AI systems. RRAO is engineered following the X-HCOME ontology engineering (OE) methodology, which provides a hybrid bottom-up (data driven), top-down (expert knowledge) OE approach, including tasks designed to assess and mitigate bias at the schema level. By developing a bias-free risk assessment ontology, our objective is to enhance the fairness of AI-driven ontology-based reoffending risk prediction systems, ultimately contributing to more fair and effective criminal justice practices.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Judicial decision making related to parole, sentencing, rehabilitation, reintegration, and public safety is often supported by the assessment of the risk of reoffending. AI prediction systems can introduce bias in the analysis of reoffending riskrelated data. Several studies criticize the fairness of such AI systems. This paper presents the Reoffending Risk Assessment Ontology (RRAO) which aims to provide a comprehensive representation of reoffending risk and recidivism knowledge integrated into ontology-based AI systems. RRAO is engineered following the X-HCOME ontology engineering (OE) methodology, which provides a hybrid bottom-up (data driven), top-down (expert knowledge) OE approach, including tasks designed to assess and mitigate bias at the schema level. By developing a bias-free risk assessment ontology, our objective is to enhance the fairness of AI-driven ontology-based reoffending risk prediction systems, ultimately contributing to more fair and effective criminal justice practices. |
Eleftherios-Efkleidis Stefanou Pavlos Bitilis, Konstantinos Kotis Current Status, Trends and Challenges in AI-Based Bradykinesia and Tremor Detection on Edge Devices Conference 16th International Conference on Information, Intelligence, Systems & Applications (IISA), 2025, ISBN: 979-8-3315-5636-5. @conference{Stefanou2025, title = {Current Status, Trends and Challenges in AI-Based Bradykinesia and Tremor Detection on Edge Devices}, author = {Eleftherios-Efkleidis Stefanou, Pavlos Bitilis, Konstantinos Kotis}, url = {https://ieeexplore.ieee.org/abstract/document/11311304}, doi = {https://doi.org/10.1109/IISA66859.2025.11311304}, isbn = {979-8-3315-5636-5}, year = {2025}, date = {2025-07-10}, booktitle = {16th International Conference on Information, Intelligence, Systems & Applications (IISA)}, pages = {1-4}, abstract = {Bradykinesia and tremor are pivotal indicators in diagnosing and managing Parkinson’s disease (PD), a progressive neurodegenerative disorder. Advances in wearable sensor technologies and AI methods have enabled real-time monitoring of these motor impairments, facilitating continuous assessment outside traditional clinical settings. This short paper focuses on recent advancements in bradykinesia and tremor detection using machine learning (ML) and deep learning (DL) techniques, while also exploring their applicability on edge devices. By leveraging inertial data, these techniques enhance the detection and analysis of movement patterns associated with PD. The paper emphasizes techniques optimized for edge deployment, which enable localized data processing, reduce latency, and enhance privacy. In addition, open-source tools and datasets are highlighted for their role in improving reproducibility and supporting system integration efforts. Finally, challenges such as variability in real-world conditions are discussed, along with opportunities for enhancing wearable-based healthcare systems through accurate and reliable motion pattern recognition on edge platforms.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Bradykinesia and tremor are pivotal indicators in diagnosing and managing Parkinson’s disease (PD), a progressive neurodegenerative disorder. Advances in wearable sensor technologies and AI methods have enabled real-time monitoring of these motor impairments, facilitating continuous assessment outside traditional clinical settings. This short paper focuses on recent advancements in bradykinesia and tremor detection using machine learning (ML) and deep learning (DL) techniques, while also exploring their applicability on edge devices. By leveraging inertial data, these techniques enhance the detection and analysis of movement patterns associated with PD. The paper emphasizes techniques optimized for edge deployment, which enable localized data processing, reduce latency, and enhance privacy. In addition, open-source tools and datasets are highlighted for their role in improving reproducibility and supporting system integration efforts. Finally, challenges such as variability in real-world conditions are discussed, along with opportunities for enhancing wearable-based healthcare systems through accurate and reliable motion pattern recognition on edge platforms. |
George Papadopoulos, George Vouros A Learning safe, constrained policies via imitation learning: Connection to Probabilistic Inference and a Naive Algorithm Journal Article arXiv, 2025. @article{Papadopoulos2025b, title = {Learning safe, constrained policies via imitation learning: Connection to Probabilistic Inference and a Naive Algorithm}, author = {George Papadopoulos, George A Vouros}, url = {https://arxiv.org/pdf/2507.06780}, doi = {https://doi.org/10.48550/arXiv.2507.06780}, year = {2025}, date = {2025-07-09}, journal = {arXiv}, abstract = {This article introduces an imitation learning method for learning maximum entropy policies that comply with constraints demonstrated by expert trajectories executing a task. The formulation of the method takes advantage of results connecting performance to bounds for the KL-divergence between demonstrated and learned policies, and its objective is rigorously justified through a connection to a probabilistic inference framework for reinforcement learning, incorporating the reinforcement learning objective and the objective to abide by constraints in an entropy maximization setting. The proposed algorithm optimizes the learning objective with dual gradient descent, supporting effective and stable training. Experiments show that the proposed method can learn effective policy models for constraints-abiding behaviour, in settings with multiple constraints of different types, accommodating different modalities of demonstrated behaviour, and with abilities to generalize.}, keywords = {}, pubstate = {published}, tppubtype = {article} } This article introduces an imitation learning method for learning maximum entropy policies that comply with constraints demonstrated by expert trajectories executing a task. The formulation of the method takes advantage of results connecting performance to bounds for the KL-divergence between demonstrated and learned policies, and its objective is rigorously justified through a connection to a probabilistic inference framework for reinforcement learning, incorporating the reinforcement learning objective and the objective to abide by constraints in an entropy maximization setting. The proposed algorithm optimizes the learning objective with dual gradient descent, supporting effective and stable training. Experiments show that the proposed method can learn effective policy models for constraints-abiding behaviour, in settings with multiple constraints of different types, accommodating different modalities of demonstrated behaviour, and with abilities to generalize. |
Piyabhum Chaysri Theodoros Tranos, George Papadopoulos George Vouros Konstantinos Blekas A Efficient Autonomous Marine Vessel Navigation with Safe Deep Reinforcement Learning Conference 2025 Symposium on Maritime Informatics and Robotics (MARIS), 2025. @conference{Chaysri2025, title = { Efficient Autonomous Marine Vessel Navigation with Safe Deep Reinforcement Learning}, author = {Piyabhum Chaysri, Theodoros Tranos, George Papadopoulos, George A Vouros, Konstantinos Blekas}, doi = {https://doi.org/10.1109/MARIS64137.2025.11139786}, year = {2025}, date = {2025-06-26}, booktitle = {2025 Symposium on Maritime Informatics and Robotics (MARIS)}, abstract = {The rise of automation and self-driving systems brings a strong focus on safety-centric decision-making, especially in complex environments with large degree of uncertainty where unpredictable interactions occur at high frequency. In this study we address the challenge of safe and efficient maritime navigation by proposing a safe Deep Reinforcement Learning scheme for training Unmanned Surface Vehicle (USV) agents. Our approach leverages the Lagrangian relaxation framework to effectively handle safety constraints, ensuring that the learned navigation policies balance goal achievement with obstacle avoidance. We address each type of static and moving obstacle separately with the aim of achieving more effective management of their impact on safe navigation. This enables the design of a more advanced constraint-aware optimization framework, enhancing USV’s ability to navigate complex maritime environment, adapt to changing traffic conditions and maintain a minimal risk of collision. Experiments were conducted in a simulated environment tailored to match realistic weather and traffic density conditions. The simulation results highlight the potential of the proposed method in developing advanced USV navigation policies that achieve high accuracy and enhanced safety.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } The rise of automation and self-driving systems brings a strong focus on safety-centric decision-making, especially in complex environments with large degree of uncertainty where unpredictable interactions occur at high frequency. In this study we address the challenge of safe and efficient maritime navigation by proposing a safe Deep Reinforcement Learning scheme for training Unmanned Surface Vehicle (USV) agents. Our approach leverages the Lagrangian relaxation framework to effectively handle safety constraints, ensuring that the learned navigation policies balance goal achievement with obstacle avoidance. We address each type of static and moving obstacle separately with the aim of achieving more effective management of their impact on safe navigation. This enables the design of a more advanced constraint-aware optimization framework, enhancing USV’s ability to navigate complex maritime environment, adapt to changing traffic conditions and maintain a minimal risk of collision. Experiments were conducted in a simulated environment tailored to match realistic weather and traffic density conditions. The simulation results highlight the potential of the proposed method in developing advanced USV navigation policies that achieve high accuracy and enhanced safety. |
Dimitrios Doumanas Efthalia Ntalouka, Costas Vassilakis Manolis Wallace Konstantinos Kotis Stitching History into Semantics: LLM-Supported Knowledge Graph Engineering for 19th-Century Greek Bookbinding Journal Article Machine Learning and Knowledge Extraction, 7 (3), pp. 59, 2025, ISSN: 2504-4990. @article{Doumanas2025c, title = {Stitching History into Semantics: LLM-Supported Knowledge Graph Engineering for 19th-Century Greek Bookbinding}, author = {Dimitrios Doumanas, Efthalia Ntalouka, Costas Vassilakis, Manolis Wallace, Konstantinos Kotis}, url = {https://www.mdpi.com/2504-4990/7/3/59}, doi = {https://doi.org/10.3390/make7030059}, issn = {2504-4990}, year = {2025}, date = {2025-06-24}, journal = {Machine Learning and Knowledge Extraction}, volume = {7}, number = {3}, pages = {59}, abstract = {Preserving cultural heritage can be efficiently supported by structured and semantic representation of historical artifacts. Bookbinding, a critical aspect of book history, provides valuable insights into past craftsmanship, material use, and conservation practices. However, existing bibliographic records often lack the depth needed to analyze bookbinding techniques, provenance, and preservation status. This paper presents a proof-of-concept system that explores how Large Language Models (LLMs) can support knowledge graph engineering within the context of 19th-century Greek bookbinding (1830–1900), and as a result, generate a domain-specific ontology and a knowledge graph. Our ontology encapsulates materials, binding techniques, artistic styles, and conservation history, integrating metadata standards like MARC and Dublin Core to ensure interoperability with existing library and archival systems. To validate its effectiveness, we construct a Neo4j knowledge graph, based on the generated ontology and utilize Cypher Queries—including LLM-generated queries—to extract insights about bookbinding practices and trends. This study also explores how semantic reasoning over the knowledge graph can identify historical binding patterns, assess book conservation needs, and infer relationships between bookbinding workshops. Unlike previous bibliographic ontologies, our approach provides a comprehensive, semantically rich representation of bookbinding history, methods and techniques, supporting scholars, conservators, and cultural heritage institutions. By demonstrating how LLMs can assist in ontology/KG creation and query generation, we introduce and evaluate a semi-automated pipeline as a methodological demonstration for studying historical bookbinding, contributing to digital humanities, book conservation, and cultural informatics. Finally, the proposed approach can be used in other domains, thus, being generally applicable in knowledge engineering.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Preserving cultural heritage can be efficiently supported by structured and semantic representation of historical artifacts. Bookbinding, a critical aspect of book history, provides valuable insights into past craftsmanship, material use, and conservation practices. However, existing bibliographic records often lack the depth needed to analyze bookbinding techniques, provenance, and preservation status. This paper presents a proof-of-concept system that explores how Large Language Models (LLMs) can support knowledge graph engineering within the context of 19th-century Greek bookbinding (1830–1900), and as a result, generate a domain-specific ontology and a knowledge graph. Our ontology encapsulates materials, binding techniques, artistic styles, and conservation history, integrating metadata standards like MARC and Dublin Core to ensure interoperability with existing library and archival systems. To validate its effectiveness, we construct a Neo4j knowledge graph, based on the generated ontology and utilize Cypher Queries—including LLM-generated queries—to extract insights about bookbinding practices and trends. This study also explores how semantic reasoning over the knowledge graph can identify historical binding patterns, assess book conservation needs, and infer relationships between bookbinding workshops. Unlike previous bibliographic ontologies, our approach provides a comprehensive, semantically rich representation of bookbinding history, methods and techniques, supporting scholars, conservators, and cultural heritage institutions. By demonstrating how LLMs can assist in ontology/KG creation and query generation, we introduce and evaluate a semi-automated pipeline as a methodological demonstration for studying historical bookbinding, contributing to digital humanities, book conservation, and cultural informatics. Finally, the proposed approach can be used in other domains, thus, being generally applicable in knowledge engineering. |
Myrto Stogia Asimina Dimara, Alexios Papaioannou Christos-Nikolaos Anagnostopoulos Konstantinos Kotis Stelios Krinidis The Role of IoT and 3D Modeling in Shaping Industry 5.0 Conference IFIP International Conference on Artificial Intelligence Applications and Innovations, 2025, ISBN: 978-3-031-97313-0. @conference{Stogia2025, title = {The Role of IoT and 3D Modeling in Shaping Industry 5.0}, author = {Myrto Stogia, Asimina Dimara, Alexios Papaioannou, Christos-Nikolaos Anagnostopoulos, Konstantinos Kotis, Stelios Krinidis}, url = {https://link.springer.com/chapter/10.1007/978-3-031-97313-0_27}, doi = {https://doi.org/10.1007/978-3-031-97313-0_27}, isbn = {978-3-031-97313-0}, year = {2025}, date = {2025-06-23}, booktitle = {IFIP International Conference on Artificial Intelligence Applications and Innovations}, pages = {353-366}, abstract = {The shift from Industry 4.0 to Industry 5.0 represents a significant transformation in industrial ecosystems, prioritizing human-machine collaboration, sustainability, and ethical Artificial Intelligence (AI). This paper provides a concise overview of the crucial role played by the Internet of Things (IoT) in advancing Digital Twin (DT) technology, particularly in improving three-dimensional modeling capabilities. IoT-driven DTs facilitate adaptive, efficient, and sustainable industrial operations by integrating real-time data, utilizing predictive analytics, and supporting smart manufacturing. Unlike Industry 4.0, which focuses on automation and cyber-physical systems, Industry 5.0 reintroduces human intelligence to ensure that technological progress aligns with ethical, social, and environmental considerations. This survey examines challenges such as scalability, interoperability, energy efficiency, and cybersecurity while exploring innovations like cognitive DTs, 5G-powered IoT networks, and AI-driven decision-making. Additionally, it highlights key technological advancements, including edge computing, neuro-symbolic and conversational AI, blockchain for secure data management, and eco-friendly IoT solutions, paving the way for a resilient, human-centric industrial future.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } The shift from Industry 4.0 to Industry 5.0 represents a significant transformation in industrial ecosystems, prioritizing human-machine collaboration, sustainability, and ethical Artificial Intelligence (AI). This paper provides a concise overview of the crucial role played by the Internet of Things (IoT) in advancing Digital Twin (DT) technology, particularly in improving three-dimensional modeling capabilities. IoT-driven DTs facilitate adaptive, efficient, and sustainable industrial operations by integrating real-time data, utilizing predictive analytics, and supporting smart manufacturing. Unlike Industry 4.0, which focuses on automation and cyber-physical systems, Industry 5.0 reintroduces human intelligence to ensure that technological progress aligns with ethical, social, and environmental considerations. This survey examines challenges such as scalability, interoperability, energy efficiency, and cybersecurity while exploring innovations like cognitive DTs, 5G-powered IoT networks, and AI-driven decision-making. Additionally, it highlights key technological advancements, including edge computing, neuro-symbolic and conversational AI, blockchain for secure data management, and eco-friendly IoT solutions, paving the way for a resilient, human-centric industrial future. |
George Giannakopoulos Andreas Sideras, Konstantinos Stamatakis Nikolaos Melanitis NAVMAT: An AI-supported naval failures knowledge management system Journal Article Expert Systems with Applications, 277 , pp. 127117, 2025. @article{Giannakopoulos2025, title = {NAVMAT: An AI-supported naval failures knowledge management system}, author = {George Giannakopoulos, Andreas Sideras, Konstantinos Stamatakis, Nikolaos Melanitis}, doi = {https://doi.org/10.1016/j.eswa.2025.127117}, year = {2025}, date = {2025-06-05}, journal = {Expert Systems with Applications}, volume = {277}, pages = {127117}, abstract = {We present “NAVMAT”, an intelligent, multilingual knowledge management platform designed to record and categorize material failure incidents reported in naval operations. This paper provides an overview of the platform, identifying its key software components and highlighting the information retrieval approach used to support user workflows. The platform primarily facilitates real-time, multilingual search and intelligent indexing, streamlining the incident management process while offering valuable insights from past incidents and knowledge resources. To achieve this, it employs a customized natural language processing pipeline integrated with a carefully engineered ontology. The ontology, regularly updated by domain experts, enriches the retrieval mechanism by instilling domain specific knowledge. This approach aims to reduce the significant variability in specialized terminology by promoting convergence towards a unified vocabulary.}, keywords = {}, pubstate = {published}, tppubtype = {article} } We present “NAVMAT”, an intelligent, multilingual knowledge management platform designed to record and categorize material failure incidents reported in naval operations. This paper provides an overview of the platform, identifying its key software components and highlighting the information retrieval approach used to support user workflows. The platform primarily facilitates real-time, multilingual search and intelligent indexing, streamlining the incident management process while offering valuable insights from past incidents and knowledge resources. To achieve this, it employs a customized natural language processing pipeline integrated with a carefully engineered ontology. The ontology, regularly updated by domain experts, enriches the retrieval mechanism by instilling domain specific knowledge. This approach aims to reduce the significant variability in specialized terminology by promoting convergence towards a unified vocabulary. |
Foteini Oikonomou Eleftherios Bailis, Sotiris Bentos Stamatis Chatzistamatis Marianna Tzortzi Konstantinos Kotis Stamatis Spirou George Tsekouras E Towards Fair Recidivism Prediction: Addressing Bias in Machine Learning for the Greek Prison System Conference 5th International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET), 2025, ISBN: 979-8-3315-3297-0. @conference{Oikonomou2025, title = {Towards Fair Recidivism Prediction: Addressing Bias in Machine Learning for the Greek Prison System}, author = {Foteini Oikonomou, Eleftherios Bailis, Sotiris Bentos, Stamatis Chatzistamatis, Marianna Tzortzi, Konstantinos Kotis, Stamatis Spirou, George E Tsekouras}, url = {https://ieeexplore.ieee.org/abstract/document/11008007}, doi = {https://doi.org/10.1109/IRASET64571.2025.11008007}, isbn = {979-8-3315-3297-0}, year = {2025}, date = {2025-05-15}, booktitle = {5th International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET)}, pages = {1-7}, abstract = {Recidivism prediction has become an essential tool in criminal justice systems, aiding decision-making in areas such as sentencing, parole, and rehabilitation. Machine learning (ML) algorithms have been widely employed to improve the accuracy of recidivism risk assessments. However, concerns about fairness and algorithmic bias have been raised, particularly in high-stakes applications. This study focuses on the Greek prison system, utilizing a dataset from Greek prisons to analyze and mitigate biases in ML-based recidivism predictions. The study primarily investigates the impact of age as a sensitive attribute and employs fairness-aware optimization techniques to reduce disparities in predictive outcomes. By incorporating fairness constraints into the training process, we demonstrate that balancing fairness and accuracy is possible. The results indicate that implementing fairness-aware ML models can significantly reduce bias, particularly against younger offenders, while maintaining acceptable predictive performance. Our findings contribute to ongoing discussions on the ethical application of AI in criminal justice and highlight the necessity of fairness-aware methodologies for equitable decision-making.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Recidivism prediction has become an essential tool in criminal justice systems, aiding decision-making in areas such as sentencing, parole, and rehabilitation. Machine learning (ML) algorithms have been widely employed to improve the accuracy of recidivism risk assessments. However, concerns about fairness and algorithmic bias have been raised, particularly in high-stakes applications. This study focuses on the Greek prison system, utilizing a dataset from Greek prisons to analyze and mitigate biases in ML-based recidivism predictions. The study primarily investigates the impact of age as a sensitive attribute and employs fairness-aware optimization techniques to reduce disparities in predictive outcomes. By incorporating fairness constraints into the training process, we demonstrate that balancing fairness and accuracy is possible. The results indicate that implementing fairness-aware ML models can significantly reduce bias, particularly against younger offenders, while maintaining acceptable predictive performance. Our findings contribute to ongoing discussions on the ethical application of AI in criminal justice and highlight the necessity of fairness-aware methodologies for equitable decision-making. |
Andreas Kontogiannis Konstantinos Papathanasiou, Yi Shen Giorgos Stamou Michael Zavlanos George Vouros M Enhancing cooperative multi-agent reinforcement learning with state modelling and adversarial exploration Journal Article arXiv, 2025. @article{Kontogiannis2025, title = {Enhancing cooperative multi-agent reinforcement learning with state modelling and adversarial exploration}, author = {Andreas Kontogiannis, Konstantinos Papathanasiou, Yi Shen, Giorgos Stamou, Michael M Zavlanos, George Vouros}, url = {https://arxiv.org/pdf/2505.05262}, doi = {https://doi.org/10.48550/arXiv.2505.05262}, year = {2025}, date = {2025-05-08}, journal = {arXiv}, abstract = {Learning to cooperate in distributed partially observable environments with no communication abilities poses significant challenges for multi-agent deep reinforcement learning (MARL). This paper addresses key concerns in this domain, focusing on inferring state representations from individual agent observations and leveraging these representations to enhance agents’ exploration and collaborative task execution policies. To this end, we propose a novel state modelling framework for cooperative MARL, where agents infer meaningful belief representations of the non-observable state, with respect to optimizing their own policies, while filtering redundant and less informative joint state information. Building upon this framework, we propose the MARL SMPE algorithm. In SMPE, agents enhance their own policy’s discriminative abilities under partial observability, explicitly by incorporating their beliefs into the policy network, and implicitly by adopting an adversarial type of exploration policies which encourages agents to discover novel, high-value states while improving the discriminative abilities of others. Experimentally, we show that SMPE outperforms state-of-the-art MARL algorithms in complex fully cooperative tasks from the MPE, LBF, and RWARE benchmarks.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Learning to cooperate in distributed partially observable environments with no communication abilities poses significant challenges for multi-agent deep reinforcement learning (MARL). This paper addresses key concerns in this domain, focusing on inferring state representations from individual agent observations and leveraging these representations to enhance agents’ exploration and collaborative task execution policies. To this end, we propose a novel state modelling framework for cooperative MARL, where agents infer meaningful belief representations of the non-observable state, with respect to optimizing their own policies, while filtering redundant and less informative joint state information. Building upon this framework, we propose the MARL SMPE algorithm. In SMPE, agents enhance their own policy’s discriminative abilities under partial observability, explicitly by incorporating their beliefs into the policy network, and implicitly by adopting an adversarial type of exploration policies which encourages agents to discover novel, high-value states while improving the discriminative abilities of others. Experimentally, we show that SMPE outperforms state-of-the-art MARL algorithms in complex fully cooperative tasks from the MPE, LBF, and RWARE benchmarks. |
Dimitris Kostadimas Vlasios Kasapakis, Konstantinos Kotis A systematic review on the combination of VR, IoT and AI technologies, and their integration in applications Journal Article Future Internet, 17 (4), pp. 163, 2025, ISSN: 1999-5903. @article{Kostadimas2025, title = {A systematic review on the combination of VR, IoT and AI technologies, and their integration in applications}, author = {Dimitris Kostadimas, Vlasios Kasapakis, Konstantinos Kotis}, url = {https://www.mdpi.com/1999-5903/17/4/163}, doi = {https://doi.org/10.3390/fi17040163}, issn = {1999-5903}, year = {2025}, date = {2025-04-07}, journal = {Future Internet}, volume = {17}, number = {4}, pages = {163}, abstract = {The convergence of Virtual Reality (VR), Artificial Intelligence (AI), and the Internet of Things (IoT) offers transformative potential across numerous sectors. However, existing studies often examine these technologies independently or in limited pairings, which overlooks the synergistic possibilities of their combined usage. This systematic review adheres to the PRISMA guidelines in order to critically analyze peer-reviewed literature from highly recognized academic databases related to the intersection of VR, AI, and IoT, and identify application domains, methodologies, tools, and key challenges. By focusing on real-life implementations and working prototypes, this review highlights state-of-the-art advancements and uncovers gaps that hinder practical adoption, such as data collection issues, interoperability barriers, and user experience challenges. The findings reveal that digital twins (DTs), AIoT systems, and immersive XR environments are promising as emerging technologies (ET), but require further development to achieve scalability and real-world impact, while in certain fields a limited amount of research is conducted until now. This review bridges theory and practice, providing a targeted foundation for future interdisciplinary research aimed at advancing practical, scalable solutions across domains such as healthcare, smart cities, industry, education, cultural heritage, and beyond. The study found that the integration of VR, AI, and IoT holds significant potential across various domains, with DTs, IoT systems, and immersive XR environments showing promising applications, but challenges such as data interoperability, user experience limitations, and scalability barriers hinder widespread adoption.}, keywords = {}, pubstate = {published}, tppubtype = {article} } The convergence of Virtual Reality (VR), Artificial Intelligence (AI), and the Internet of Things (IoT) offers transformative potential across numerous sectors. However, existing studies often examine these technologies independently or in limited pairings, which overlooks the synergistic possibilities of their combined usage. This systematic review adheres to the PRISMA guidelines in order to critically analyze peer-reviewed literature from highly recognized academic databases related to the intersection of VR, AI, and IoT, and identify application domains, methodologies, tools, and key challenges. By focusing on real-life implementations and working prototypes, this review highlights state-of-the-art advancements and uncovers gaps that hinder practical adoption, such as data collection issues, interoperability barriers, and user experience challenges. The findings reveal that digital twins (DTs), AIoT systems, and immersive XR environments are promising as emerging technologies (ET), but require further development to achieve scalability and real-world impact, while in certain fields a limited amount of research is conducted until now. This review bridges theory and practice, providing a targeted foundation for future interdisciplinary research aimed at advancing practical, scalable solutions across domains such as healthcare, smart cities, industry, education, cultural heritage, and beyond. The study found that the integration of VR, AI, and IoT holds significant potential across various domains, with DTs, IoT systems, and immersive XR environments showing promising applications, but challenges such as data interoperability, user experience limitations, and scalability barriers hinder widespread adoption. |
Georgios Bouchouras Georgios Sofianidis, Konstantinos Kotis Predicting freezing of gait in parkinson’s disease: A machine-learning-based approach in on and off medication states Journal Article Journal of Clinical Medicine, 14 (6), pp. 2120, 2025, ISSN: 2077-0383. @article{Bouchouras2025c, title = {Predicting freezing of gait in parkinson’s disease: A machine-learning-based approach in on and off medication states}, author = {Georgios Bouchouras, Georgios Sofianidis, Konstantinos Kotis}, url = {https://www.mdpi.com/2077-0383/14/6/2120}, doi = {https://doi.org/10.3390/jcm14062120}, issn = {2077-0383}, year = {2025}, date = {2025-03-20}, journal = {Journal of Clinical Medicine}, volume = {14}, number = {6}, pages = {2120}, abstract = {Freezing of gait (FoG) is a debilitating motor symptom of Parkinson’s disease (PD), characterized by sudden episodes where patients struggle to initiate or sustain movement, often describing a sensation of their feet being “glued to the ground.” This study investigates the potential of machine-learning (ML) models to predict FoG severity in PD patients, focusing on the influence of dopaminergic medication by comparing gait parameters in ON and OFF medication states. Methods: Specifically, this study employed spatiotemporal gait features to develop a predictive model for FoG severity, leveraging a random forest regressor to identify the most influential gait parameters associated with this in each medication state. The results indicate that the model achieved higher predictive performance in the OFF-medication condition (R² = 0.82, MAE = 2.25, MSE = 15.23) compared to the ON-medication condition (R² = 0.52, MAE = 4.16, MSE = 42.00). Results: These findings suggest that dopaminergic treatment alters gait dynamics, potentially reducing the reliability of FoG predictions when patients are medicated. Feature importance analysis revealed distinct gait characteristics associated with FoG severity across medication states. In the OFF condition, step length parameters, particularly left step length mean, were the most dominant predictors, alongside swing time and stride width, indicating the role of spatial and temporal gait control in FoG severity without medication. In contrast, under the ON medication condition, stride width and gait speed emerged as the most influential predictors, followed by stepping frequency, reflecting how medication influences stability and movement rhythm. Conclusions: These findings highlight the need for predictive models that account for medication-induced gait variability, ensuring more reliable FoG detection. By integrating spatiotemporal gait analysis and ML-based prediction, this study contributes to the development of personalized intervention strategies for PD patients experiencing FoG episodes.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Freezing of gait (FoG) is a debilitating motor symptom of Parkinson’s disease (PD), characterized by sudden episodes where patients struggle to initiate or sustain movement, often describing a sensation of their feet being “glued to the ground.” This study investigates the potential of machine-learning (ML) models to predict FoG severity in PD patients, focusing on the influence of dopaminergic medication by comparing gait parameters in ON and OFF medication states. Methods: Specifically, this study employed spatiotemporal gait features to develop a predictive model for FoG severity, leveraging a random forest regressor to identify the most influential gait parameters associated with this in each medication state. The results indicate that the model achieved higher predictive performance in the OFF-medication condition (R² = 0.82, MAE = 2.25, MSE = 15.23) compared to the ON-medication condition (R² = 0.52, MAE = 4.16, MSE = 42.00). Results: These findings suggest that dopaminergic treatment alters gait dynamics, potentially reducing the reliability of FoG predictions when patients are medicated. Feature importance analysis revealed distinct gait characteristics associated with FoG severity across medication states. In the OFF condition, step length parameters, particularly left step length mean, were the most dominant predictors, alongside swing time and stride width, indicating the role of spatial and temporal gait control in FoG severity without medication. In contrast, under the ON medication condition, stride width and gait speed emerged as the most influential predictors, followed by stepping frequency, reflecting how medication influences stability and movement rhythm. Conclusions: These findings highlight the need for predictive models that account for medication-induced gait variability, ensuring more reliable FoG detection. By integrating spatiotemporal gait analysis and ML-based prediction, this study contributes to the development of personalized intervention strategies for PD patients experiencing FoG episodes. |
Despoina P Kiouri Georgios C Batsis, Thomas Mavromoustakos Alessandro Giuliani Christos Chasapis T Structure-Based Modeling of the Gut Bacteria–Host Interactome Through Statistical Analysis of Domain–Domain Associations Using Machine Learning Journal Article BioTech, 14 (1), pp. 13, 2025. @article{Kiouri2025c, title = {Structure-Based Modeling of the Gut Bacteria–Host Interactome Through Statistical Analysis of Domain–Domain Associations Using Machine Learning}, author = {Despoina P Kiouri, Georgios C Batsis, Thomas Mavromoustakos, Alessandro Giuliani, Christos T Chasapis}, url = {https://www.mdpi.com/2673-6284/14/1/13/pdf?version=1740479495}, doi = {https://doi.org/10.3390/biotech14010013}, year = {2025}, date = {2025-02-25}, journal = {BioTech}, volume = {14}, number = {1}, pages = {13}, abstract = {The gut microbiome, a complex ecosystem of microorganisms, plays a pivotal role in human health and disease. The gut microbiome’s influence extends beyond the digestive system to various organs, and its imbalance is linked to a wide range of diseases, including cancer and neurodevelopmental, inflammatory, metabolic, cardiovascular, autoimmune, and psychiatric diseases. Despite its significance, the interactions between gut bacteria and human proteins remain understudied, with less than 20,000 experimentally validated protein interactions between the host and any bacteria species. This study addresses this knowledge gap by predicting a protein–protein interaction network between gut bacterial and human proteins. Using statistical associations between Pfam domains, a comprehensive dataset of over one million experimentally validated pan-bacterial–human protein interactions, as well as inter- and intra-species protein interactions from various organisms, were used for the development of a machine learning-based prediction method to uncover key regulatory molecules in this dynamic system. This study’s findings contribute to the understanding of the intricate gut microbiome–host relationship and pave the way for future experimental validation and therapeutic strategies targeting the gut microbiome interplay.}, keywords = {}, pubstate = {published}, tppubtype = {article} } The gut microbiome, a complex ecosystem of microorganisms, plays a pivotal role in human health and disease. The gut microbiome’s influence extends beyond the digestive system to various organs, and its imbalance is linked to a wide range of diseases, including cancer and neurodevelopmental, inflammatory, metabolic, cardiovascular, autoimmune, and psychiatric diseases. Despite its significance, the interactions between gut bacteria and human proteins remain understudied, with less than 20,000 experimentally validated protein interactions between the host and any bacteria species. This study addresses this knowledge gap by predicting a protein–protein interaction network between gut bacterial and human proteins. Using statistical associations between Pfam domains, a comprehensive dataset of over one million experimentally validated pan-bacterial–human protein interactions, as well as inter- and intra-species protein interactions from various organisms, were used for the development of a machine learning-based prediction method to uncover key regulatory molecules in this dynamic system. This study’s findings contribute to the understanding of the intricate gut microbiome–host relationship and pave the way for future experimental validation and therapeutic strategies targeting the gut microbiome interplay. |
Natalia Koliou, George Vouros Ranking Joint Policies in Dynamic Games using Evolutionary Dynamics Journal Article arXiv, 2025. @article{Koliou2025, title = {Ranking Joint Policies in Dynamic Games using Evolutionary Dynamics}, author = {Natalia Koliou, George Vouros}, url = {https://arxiv.org/pdf/2502.14724}, doi = {https://doi.org/10.48550/arXiv.2502.14724}, year = {2025}, date = {2025-02-20}, journal = {arXiv}, abstract = {Game-theoretic solution concepts, such as the Nash equilibrium, have been key to finding stable joint actions in multi-player games. However, it has been shown that the dynamics of agents’ interactions, even in simple two-player games with few strategies, are incapable of reaching Nash equilibria, exhibiting complex and unpredictable behavior. Instead, evolutionary approaches can describe the long-term persistence of strategies and filter out transient ones, accounting for the long-term dynamics of agents’ interactions. Our goal is to identify agents’ joint strategies that result in stable behavior, being resistant to changes, while also accounting for agents’ payoffs, in dynamic games. Towards this goal, and building on previous results, this paper proposes transforming dynamic games into their empirical forms by considering agents’ strategies instead of agents’ actions, and applying the evolutionary methodology -Rank to evaluate and rank strategy profiles according to their long-term dynamics. This methodology not only allows us to identify joint strategies that are strong through agents’ long-term interactions, but also provides a descriptive, transparent framework regarding the high ranking of these strategies. Experiments report on agents that aim to collaboratively solve a stochastic version of the graph coloring problem. We consider different styles of play as strategies to define the empirical game, and train policies realizing these strategies, using the DQN algorithm. Then we run simulations to generate the payoff matrix required by -Rank to rank joint strategies.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Game-theoretic solution concepts, such as the Nash equilibrium, have been key to finding stable joint actions in multi-player games. However, it has been shown that the dynamics of agents’ interactions, even in simple two-player games with few strategies, are incapable of reaching Nash equilibria, exhibiting complex and unpredictable behavior. Instead, evolutionary approaches can describe the long-term persistence of strategies and filter out transient ones, accounting for the long-term dynamics of agents’ interactions. Our goal is to identify agents’ joint strategies that result in stable behavior, being resistant to changes, while also accounting for agents’ payoffs, in dynamic games. Towards this goal, and building on previous results, this paper proposes transforming dynamic games into their empirical forms by considering agents’ strategies instead of agents’ actions, and applying the evolutionary methodology -Rank to evaluate and rank strategy profiles according to their long-term dynamics. This methodology not only allows us to identify joint strategies that are strong through agents’ long-term interactions, but also provides a descriptive, transparent framework regarding the high ranking of these strategies. Experiments report on agents that aim to collaboratively solve a stochastic version of the graph coloring problem. We consider different styles of play as strategies to define the empirical game, and train policies realizing these strategies, using the DQN algorithm. Then we run simulations to generate the payoff matrix required by -Rank to rank joint strategies. |
Dimitrios Doumanas Andreas Soularidis, Dimitris Spiliotopoulos Costas Vassilakis Konstantinos Kotis Fine-tuning large language models for ontology engineering: A comparative analysis of GPT-4 and Mistral Journal Article Applied Sciences, 15 (4), pp. 2146, 2025, ISSN: 2076-3417. @article{Doumanas2025b, title = {Fine-tuning large language models for ontology engineering: A comparative analysis of GPT-4 and Mistral}, author = {Dimitrios Doumanas, Andreas Soularidis, Dimitris Spiliotopoulos, Costas Vassilakis, Konstantinos Kotis}, doi = {https://doi.org/10.3390/app15042146}, issn = {2076-3417}, year = {2025}, date = {2025-02-18}, journal = {Applied Sciences}, volume = {15}, number = {4}, pages = {2146}, abstract = {Ontology engineering (OE) plays a critical role in modeling and managing structured knowledge across various domains. This study examines the performance of fine-tuned large language models (LLMs), specifically GPT-4 and Mistral 7B, in efficiently automating OE tasks. Foundational OE textbooks are used as the basis for dataset creation and for feeding the LLMs. The methodology involved segmenting texts into manageable chapters, generating question–answer pairs, and translating visual elements into description logic to curate fine-tuned datasets in JSONL format. This research aims to enhance the models’ abilities to generate domain-specific ontologies, with hypotheses asserting that fine-tuned LLMs would outperform base models, and that domain-specific datasets would significantly improve their performance. Comparative experiments revealed that GPT-4 demonstrated superior accuracy and adherence to ontology syntax, albeit with higher computational costs. Conversely, Mistral 7B excelled in speed and cost efficiency but struggled with domain-specific tasks, often generating outputs that lacked syntactical precision and relevance. The presented results highlight the necessity of integrating domain-specific datasets to improve contextual understanding and practical utility in specialized applications, such as Search and Rescue (SAR) missions in wildfire incidents. Both models, despite their limitations, exhibited potential in understanding OE principles. However, their performance underscored the importance of aligning training data with domain-specific knowledge to emulate human expertise effectively. This study, based on and extending our previous work on the topic, concludes that fine-tuned LLMs with targeted datasets enhance their utility in OE, offering insights into improving future models for domain-specific applications. The findings advocate further exploration of hybrid solutions to balance accuracy and efficiency.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Ontology engineering (OE) plays a critical role in modeling and managing structured knowledge across various domains. This study examines the performance of fine-tuned large language models (LLMs), specifically GPT-4 and Mistral 7B, in efficiently automating OE tasks. Foundational OE textbooks are used as the basis for dataset creation and for feeding the LLMs. The methodology involved segmenting texts into manageable chapters, generating question–answer pairs, and translating visual elements into description logic to curate fine-tuned datasets in JSONL format. This research aims to enhance the models’ abilities to generate domain-specific ontologies, with hypotheses asserting that fine-tuned LLMs would outperform base models, and that domain-specific datasets would significantly improve their performance. Comparative experiments revealed that GPT-4 demonstrated superior accuracy and adherence to ontology syntax, albeit with higher computational costs. Conversely, Mistral 7B excelled in speed and cost efficiency but struggled with domain-specific tasks, often generating outputs that lacked syntactical precision and relevance. The presented results highlight the necessity of integrating domain-specific datasets to improve contextual understanding and practical utility in specialized applications, such as Search and Rescue (SAR) missions in wildfire incidents. Both models, despite their limitations, exhibited potential in understanding OE principles. However, their performance underscored the importance of aligning training data with domain-specific knowledge to emulate human expertise effectively. This study, based on and extending our previous work on the topic, concludes that fine-tuned LLMs with targeted datasets enhance their utility in OE, offering insights into improving future models for domain-specific applications. The findings advocate further exploration of hybrid solutions to balance accuracy and efficiency. |
Christos Spatharis Konstantinos Blekas, George Santipantakis George Vouros Modular and Multimodal Generative Adversarial Imitation Learning for Modeling Flight Trajectories Journal Article Journal of Air Transportation, 33 (3), pp. 188-204, 2025. @article{Spatharis2025, title = {Modular and Multimodal Generative Adversarial Imitation Learning for Modeling Flight Trajectories}, author = {Christos Spatharis, Konstantinos Blekas, George Santipantakis, George Vouros}, doi = {https://doi.org/10.2514/1.D0396}, year = {2025}, date = {2025-02-17}, journal = {Journal of Air Transportation}, volume = {33}, number = {3}, pages = {188-204}, abstract = {We aim to imitate the execution of modular tasks by exploiting unsegmented trajectories that demonstrate the execution of these tasks. This is challenging since the execution of tasks follows different modes (i.e., patterns of behavior), which may exist in various mixtures within subtasks, and the identification of trajectories’ modules (i.e., subtrajectories executing subtasks) may not be easy. This paper addresses the modularity of trajectories in conjunction with multimodality toward imitating the execution of aircraft trajectories. It proposes an imitation learning framework for the aircraft trajectory prediction problem, which segments demonstrated aircraft trajectories into subtrajectories corresponding to flight phases. This facilitates disentangling modes and learning a mixture of policies per flight phase. While trajectories are segmented using domain-specific rules, a mixture of policies per flight phase is learned by a generative multimodal imitation learning method. This modular approach enables accurate prediction of both modes and subtrajectories, which finally results in predicting the evolution of the aircraft state across the whole trajectory in a compositional way. Experiments using a real-world dataset of long flights show the potential of the proposed framework to disentangle multimodal trajectories in real-world settings and predict trajectories with high accuracy, in comparison to methods that do not exploit subtrajectories.}, keywords = {}, pubstate = {published}, tppubtype = {article} } We aim to imitate the execution of modular tasks by exploiting unsegmented trajectories that demonstrate the execution of these tasks. This is challenging since the execution of tasks follows different modes (i.e., patterns of behavior), which may exist in various mixtures within subtasks, and the identification of trajectories’ modules (i.e., subtrajectories executing subtasks) may not be easy. This paper addresses the modularity of trajectories in conjunction with multimodality toward imitating the execution of aircraft trajectories. It proposes an imitation learning framework for the aircraft trajectory prediction problem, which segments demonstrated aircraft trajectories into subtrajectories corresponding to flight phases. This facilitates disentangling modes and learning a mixture of policies per flight phase. While trajectories are segmented using domain-specific rules, a mixture of policies per flight phase is learned by a generative multimodal imitation learning method. This modular approach enables accurate prediction of both modes and subtrajectories, which finally results in predicting the evolution of the aircraft state across the whole trajectory in a compositional way. Experiments using a real-world dataset of long flights show the potential of the proposed framework to disentangle multimodal trajectories in real-world settings and predict trajectories with high accuracy, in comparison to methods that do not exploit subtrajectories. |
Despoina P Kiouri Georgios C Batsis, Christos Chasapis T Structure-Based Deep Learning Framework for Modeling Human–Gut Bacterial Protein Interactions Journal Article Proteomes, 13 (1), pp. 10, 2025. @article{Kiouri2025b, title = {Structure-Based Deep Learning Framework for Modeling Human–Gut Bacterial Protein Interactions}, author = {Despoina P Kiouri, Georgios C Batsis, Christos T Chasapis}, url = {https://www.mdpi.com/2227-7382/13/1/10/pdf?version=1739790880}, doi = {https://doi.org/10.3390/proteomes13010010}, year = {2025}, date = {2025-02-17}, journal = {Proteomes}, volume = {13}, number = {1}, pages = {10}, abstract = {The interaction network between the human host proteins and the proteins of the gut bacteria is essential for the establishment of human health, and its dysregulation directly contributes to disease development. Despite its great importance, experimental data on protein–protein interactions (PPIs) between these species are sparse due to experimental limitations. Methods: This study presents a deep learning-based framework for predicting PPIs between human and gut bacterial proteins using structural data. The framework leverages graph-based protein representations and variational autoencoders (VAEs) to extract structural embeddings from protein graphs, which are then fused through a Bi-directional Cross-Attention module to predict interactions. The model addresses common challenges in PPI datasets, such as class imbalance, using focal loss to emphasize harder-to-classify samples. Results: The results demonstrated that this framework exhibits robust performance, with high precision and recall across validation and test datasets, underscoring its generalizability. By incorporating proteoforms in the analysis, the model accounts for the structural complexity within proteomes, making predictions biologically relevant. Conclusions: These findings offer a scalable tool for investigating the interactions between the host and the gut microbiota, potentially yielding new treatment targets and diagnostics for disorders linked to the microbiome.}, keywords = {}, pubstate = {published}, tppubtype = {article} } The interaction network between the human host proteins and the proteins of the gut bacteria is essential for the establishment of human health, and its dysregulation directly contributes to disease development. Despite its great importance, experimental data on protein–protein interactions (PPIs) between these species are sparse due to experimental limitations. Methods: This study presents a deep learning-based framework for predicting PPIs between human and gut bacterial proteins using structural data. The framework leverages graph-based protein representations and variational autoencoders (VAEs) to extract structural embeddings from protein graphs, which are then fused through a Bi-directional Cross-Attention module to predict interactions. The model addresses common challenges in PPI datasets, such as class imbalance, using focal loss to emphasize harder-to-classify samples. Results: The results demonstrated that this framework exhibits robust performance, with high precision and recall across validation and test datasets, underscoring its generalizability. By incorporating proteoforms in the analysis, the model accounts for the structural complexity within proteomes, making predictions biologically relevant. Conclusions: These findings offer a scalable tool for investigating the interactions between the host and the gut microbiota, potentially yielding new treatment targets and diagnostics for disorders linked to the microbiome. |
George Papadopoulos Andreas Kontogiannis, Foteini Papadopoulou Chaido Poulianou Ioannis Koumentis George Vouros An extended benchmarking of multi-agent reinforcement learning algorithms in complex fully cooperative tasks Journal Article arXiv, 2025. @article{Papadopoulos2025, title = {An extended benchmarking of multi-agent reinforcement learning algorithms in complex fully cooperative tasks}, author = {George Papadopoulos, Andreas Kontogiannis, Foteini Papadopoulou, Chaido Poulianou, Ioannis Koumentis, George Vouros}, url = {https://arxiv.org/pdf/2502.04773}, doi = {https://doi.org/10.48550/arXiv.2502.04773}, year = {2025}, date = {2025-02-07}, journal = {arXiv}, abstract = {Multi-Agent Reinforcement Learning (MARL) has recently emerged as a significant area of research. However, MARL evaluation often lacks systematic diversity, hindering a comprehensive understanding of algorithms’ capabilities. In particular, cooperative MARL algorithms are predominantly evaluated on benchmarks such as SMAC and GRF, which primarily feature team game scenarios without assessing adequately various aspects of agents’ capabilities required in fully cooperative real-world tasks such as multi-robot cooperation and warehouse, resource management, search and rescue, and human-AI cooperation. Moreover, MARL algorithms are mainly evaluated on low dimensional state spaces, and thus their performance on high-dimensional (e.g., image) observations is not well-studied. To fill this gap, this paper highlights the crucial need for expanding systematic evaluation across a wider array of existing benchmarks. To this end, we conduct extensive evaluation and comparisons of well-known MARL algorithms on complex fully cooperative benchmarks, including tasks with images as agents’ observations. Interestingly, our analysis shows that many algorithms, hailed as state-of-the-art on SMAC and GRF, may underperform standard MARL baselines on fully cooperative benchmarks. Finally, towards more systematic and better evaluation of cooperative MARL algorithms, we have open-sourced PyMARLzoo+, an extension of the widely used (E)PyMARL libraries, which addresses an open challenge from [49], facilitating seamless integration and support with all benchmarks of PettingZoo, as well as Overcooked, PressurePlate, Capture Target and Box Pushing.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Multi-Agent Reinforcement Learning (MARL) has recently emerged as a significant area of research. However, MARL evaluation often lacks systematic diversity, hindering a comprehensive understanding of algorithms’ capabilities. In particular, cooperative MARL algorithms are predominantly evaluated on benchmarks such as SMAC and GRF, which primarily feature team game scenarios without assessing adequately various aspects of agents’ capabilities required in fully cooperative real-world tasks such as multi-robot cooperation and warehouse, resource management, search and rescue, and human-AI cooperation. Moreover, MARL algorithms are mainly evaluated on low dimensional state spaces, and thus their performance on high-dimensional (e.g., image) observations is not well-studied. To fill this gap, this paper highlights the crucial need for expanding systematic evaluation across a wider array of existing benchmarks. To this end, we conduct extensive evaluation and comparisons of well-known MARL algorithms on complex fully cooperative benchmarks, including tasks with images as agents’ observations. Interestingly, our analysis shows that many algorithms, hailed as state-of-the-art on SMAC and GRF, may underperform standard MARL baselines on fully cooperative benchmarks. Finally, towards more systematic and better evaluation of cooperative MARL algorithms, we have open-sourced PyMARLzoo+, an extension of the widely used (E)PyMARL libraries, which addresses an open challenge from [49], facilitating seamless integration and support with all benchmarks of PettingZoo, as well as Overcooked, PressurePlate, Capture Target and Box Pushing. |
Fotis Assimakopoulos Costas Vassilakis, Dionisis Margaris Konstantinos Kotis Dimitris Spiliotopoulos AI and related technologies in the fields of smart agriculture: A review Journal Article Information, 16 (2), pp. 100, 2025. @article{Assimakopoulos2025, title = {AI and related technologies in the fields of smart agriculture: A review}, author = {Fotis Assimakopoulos, Costas Vassilakis, Dionisis Margaris, Konstantinos Kotis, Dimitris Spiliotopoulos}, url = {https://www.mdpi.com/2078-2489/16/2/100/pdf?version=1738918054}, doi = {https://doi.org/10.3390/info16020100}, year = {2025}, date = {2025-02-02}, journal = {Information}, volume = {16}, number = {2}, pages = {100}, abstract = {The integration of cutting-edge technologies—such as the Internet of Things (IoT), artificial intelligence (AI), machine learning (ML), and various emerging technologies—is revolutionizing agricultural practices, enhancing productivity, sustainability, and efficiency. The objective of this study is to review the literature regarding the development and evolution of AI as well as other emerging technologies in the various fields of Agriculture as they are developed and transformed by integrating the above technologies. The areas examined in this study are open field smart farming, vertical and indoor farming, zero waste agriculture, precision livestock farming, smart greenhouses, and regenerative agriculture. This paper links current research, technological innovations, and case studies to present a comprehensive review of these emerging technologies being developed in the context of smart agriculture, for the benefit of farmers and consumers in general. By exploring practical applications and future perspectives, this work aims to provide valuable insights to address global food security challenges, minimize environmental impacts, and support sustainable development goals through the application of new technologies.}, keywords = {}, pubstate = {published}, tppubtype = {article} } The integration of cutting-edge technologies—such as the Internet of Things (IoT), artificial intelligence (AI), machine learning (ML), and various emerging technologies—is revolutionizing agricultural practices, enhancing productivity, sustainability, and efficiency. The objective of this study is to review the literature regarding the development and evolution of AI as well as other emerging technologies in the various fields of Agriculture as they are developed and transformed by integrating the above technologies. The areas examined in this study are open field smart farming, vertical and indoor farming, zero waste agriculture, precision livestock farming, smart greenhouses, and regenerative agriculture. This paper links current research, technological innovations, and case studies to present a comprehensive review of these emerging technologies being developed in the context of smart agriculture, for the benefit of farmers and consumers in general. By exploring practical applications and future perspectives, this work aims to provide valuable insights to address global food security challenges, minimize environmental impacts, and support sustainable development goals through the application of new technologies. |
Dimitrios Doumanas Georgios Bouchouras, Andreas Soularidis Konstantinos Kotis George Vouros From human-to LLM-centered collaborative ontology engineering Journal Article Applied Ontology, 19 (4), pp. 334-367, 2025. @article{Doumanas2025, title = {From human-to LLM-centered collaborative ontology engineering}, author = {Dimitrios Doumanas, Georgios Bouchouras, Andreas Soularidis, Konstantinos Kotis, George Vouros}, doi = {https://doi.org/10.1177/15705838241305067}, year = {2025}, date = {2025-01-31}, journal = {Applied Ontology}, volume = {19}, number = {4}, pages = {334-367}, abstract = {In the continuously evolving landscape of knowledge engineering, the symbiosis and teaming of humans and machines emerge as a pivotal new domain. This article explores the multifaceted realms of human and machine collaborative ontology engineering (OE). The goal of the presented work is to explore the potential of Large Language Models (LLMs) to speed up and automate the processes of collaborative OE, experimenting with different levels of LLM involvement. The proposed approach is based on a human-centered approach, that is, the HCOME approach to collaborative OE, and follows a process of exploring the declining involvement of humans and the parallel increase of LLM involvement, concluding at a level of automation where the OE is exclusively performed by LLMs. This experimentation is organized based on a series of human/LLM collaboration levels (a spectrum of OE), each one aligned to a specific OE methodology, that is, Level-0 HCOME (Human), Level-1 X-HCOME (Human and LLMs), Level-2 SimX-HCOME (LLMs and Human), and Level-3 Sim-HCOME (LLMs). The evaluation of these methodologies (one per level) is performed by measuring the similarity of the generated ontologies against “reference” ontologies (precision, recall, and F1-score of reference-to-LLM-generated ontological mappings). The results presented in this paper demonstrate that while LLMs significantly expedite the OE process, the accuracy and completeness of the resulting ontologies are notably enhanced by maintaining a high level of human involvement. This study is expected to contribute to a deeper understanding of evolving dynamics in LLM-based/enhanced OE, paving the way for future advancements toward more effective collaborative OE frameworks.}, keywords = {}, pubstate = {published}, tppubtype = {article} } In the continuously evolving landscape of knowledge engineering, the symbiosis and teaming of humans and machines emerge as a pivotal new domain. This article explores the multifaceted realms of human and machine collaborative ontology engineering (OE). The goal of the presented work is to explore the potential of Large Language Models (LLMs) to speed up and automate the processes of collaborative OE, experimenting with different levels of LLM involvement. The proposed approach is based on a human-centered approach, that is, the HCOME approach to collaborative OE, and follows a process of exploring the declining involvement of humans and the parallel increase of LLM involvement, concluding at a level of automation where the OE is exclusively performed by LLMs. This experimentation is organized based on a series of human/LLM collaboration levels (a spectrum of OE), each one aligned to a specific OE methodology, that is, Level-0 HCOME (Human), Level-1 X-HCOME (Human and LLMs), Level-2 SimX-HCOME (LLMs and Human), and Level-3 Sim-HCOME (LLMs). The evaluation of these methodologies (one per level) is performed by measuring the similarity of the generated ontologies against “reference” ontologies (precision, recall, and F1-score of reference-to-LLM-generated ontological mappings). The results presented in this paper demonstrate that while LLMs significantly expedite the OE process, the accuracy and completeness of the resulting ontologies are notably enhanced by maintaining a high level of human involvement. This study is expected to contribute to a deeper understanding of evolving dynamics in LLM-based/enhanced OE, paving the way for future advancements toward more effective collaborative OE frameworks. |
Despoina P Kiouri Georgios C Batsis, Christos Chasapis T Structure-based approaches for protein–protein interaction prediction using machine learning and deep learning Journal Article Biomolecules, 15 (1), pp. 141, 2025. @article{Kiouri2025, title = {Structure-based approaches for protein–protein interaction prediction using machine learning and deep learning}, author = {Despoina P Kiouri, Georgios C Batsis, Christos T Chasapis}, url = {https://www.mdpi.com/2218-273X/15/1/141/pdf?version=1737097468}, doi = {https://doi.org/10.3390/biom15010141}, year = {2025}, date = {2025-01-17}, journal = {Biomolecules}, volume = {15}, number = {1}, pages = {141}, abstract = {Protein–Protein Interaction (PPI) prediction plays a pivotal role in understanding cellular processes and uncovering molecular mechanisms underlying health and disease. Structure-based PPI prediction has emerged as a robust alternative to sequence-based methods, offering greater biological accuracy by integrating three-dimensional spatial and biochemical features. This work summarizes the recent advances in computational approaches leveraging protein structure information for PPI prediction, focusing on machine learning (ML) and deep learning (DL) techniques. These methods not only improve predictive accuracy but also provide insights into functional sites, such as binding and catalytic residues. However, challenges such as limited high-resolution structural data and the need for effective negative sampling persist. Through the integration of experimental and computational tools, structure-based prediction paves the way for comprehensive proteomic network analysis, holding promise for advancements in drug discovery, biomarker identification, and personalized medicine. Future directions include enhancing scalability and dataset reliability to expand these approaches across diverse proteomes.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Protein–Protein Interaction (PPI) prediction plays a pivotal role in understanding cellular processes and uncovering molecular mechanisms underlying health and disease. Structure-based PPI prediction has emerged as a robust alternative to sequence-based methods, offering greater biological accuracy by integrating three-dimensional spatial and biochemical features. This work summarizes the recent advances in computational approaches leveraging protein structure information for PPI prediction, focusing on machine learning (ML) and deep learning (DL) techniques. These methods not only improve predictive accuracy but also provide insights into functional sites, such as binding and catalytic residues. However, challenges such as limited high-resolution structural data and the need for effective negative sampling persist. Through the integration of experimental and computational tools, structure-based prediction paves the way for comprehensive proteomic network analysis, holding promise for advancements in drug discovery, biomarker identification, and personalized medicine. Future directions include enhancing scalability and dataset reliability to expand these approaches across diverse proteomes. |
Georgios Bouchouras, Konstantinos Kotis Algorithms, 18 (1), pp. 34, 2025. @article{Bouchouras2025b, title = {Integrating artificial intelligence, internet of things, and sensor-based technologies: a systematic review of methodologies in autism Spectrum disorder detection}, author = {Georgios Bouchouras, Konstantinos Kotis}, url = {https://www.mdpi.com/1999-4893/18/1/34}, doi = {https://doi.org/10.3390/a18010034}, year = {2025}, date = {2025-01-09}, journal = {Algorithms}, volume = {18}, number = {1}, pages = {34}, abstract = {This paper presents a systematic review of the emerging applications of artificial intelligence (AI), Internet of Things (IoT), and sensor-based technologies in the diagnosis of autism spectrum disorder (ASD). The integration of these technologies has led to promising advances in identifying unique behavioral, physiological, and neuroanatomical markers associated with ASD. Through an examination of recent studies, we explore how technologies such as wearable sensors, eye-tracking systems, virtual reality environments, neuroimaging, and microbiome analysis contribute to a holistic approach to ASD diagnostics. The analysis reveals how these technologies facilitate non-invasive, real-time assessments across diverse settings, enhancing both diagnostic accuracy and accessibility. The findings underscore the transformative potential of AI, IoT, and sensor-based driven tools in providing personalized and continuous ASD detection, advocating for data-driven approaches that extend beyond traditional methodologies. Ultimately, this review emphasizes the role of technology in improving ASD diagnostic processes, paving the way for targeted and individualized assessments.}, keywords = {}, pubstate = {published}, tppubtype = {article} } This paper presents a systematic review of the emerging applications of artificial intelligence (AI), Internet of Things (IoT), and sensor-based technologies in the diagnosis of autism spectrum disorder (ASD). The integration of these technologies has led to promising advances in identifying unique behavioral, physiological, and neuroanatomical markers associated with ASD. Through an examination of recent studies, we explore how technologies such as wearable sensors, eye-tracking systems, virtual reality environments, neuroimaging, and microbiome analysis contribute to a holistic approach to ASD diagnostics. The analysis reveals how these technologies facilitate non-invasive, real-time assessments across diverse settings, enhancing both diagnostic accuracy and accessibility. The findings underscore the transformative potential of AI, IoT, and sensor-based driven tools in providing personalized and continuous ASD detection, advocating for data-driven approaches that extend beyond traditional methodologies. Ultimately, this review emphasizes the role of technology in improving ASD diagnostic processes, paving the way for targeted and individualized assessments. |
2024 |
Georgios Bouchouras Georgios Sofianidis, Konstantinos Kotis Temporal Anomaly Detection in Attention-Deficit/Hyperactivity Disorder Using Recurrent Neural Networks Journal Article Cureus, 16 (12), 2024. @article{Bouchouras2024b, title = {Temporal Anomaly Detection in Attention-Deficit/Hyperactivity Disorder Using Recurrent Neural Networks}, author = {Georgios Bouchouras, Georgios Sofianidis, Konstantinos Kotis}, url = {https://assets.cureus.com/uploads/original_article/pdf/324320/20250127-834463-ohpyph.pdf}, doi = {https://doi.org/10.7759/cureus.76496}, year = {2024}, date = {2024-12-27}, journal = {Cureus}, volume = {16}, number = {12}, abstract = {Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental condition marked by movement hyperactivity, often persisting into adulthood. Understanding the movement patterns associated with ADHD is crucial for improving diagnostic precision and tailoring interventions. This study leverages the HYPERAKTIV dataset, which includes high-resolution temporal data on motor activity from people diagnosed with ADHD. We used the isolation forest algorithm to detect anomalies in activity data, followed by the development of a recurrent neural network (RNN) model to predict these anomalies over time. The RNN model demonstrated high predictive accuracy, with a mean accuracy of 0.953 and a mean loss of 0.124 for participants with ADHD. These findings suggest that machine learning techniques, particularly RNNs, can effectively identify and predict anomalies in temporal motor activity data, offering objective insights into ADHD-related movement behaviors. This approach is promising for informing personalized interventions and improving clinical decision-making in the management of ADHD.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental condition marked by movement hyperactivity, often persisting into adulthood. Understanding the movement patterns associated with ADHD is crucial for improving diagnostic precision and tailoring interventions. This study leverages the HYPERAKTIV dataset, which includes high-resolution temporal data on motor activity from people diagnosed with ADHD. We used the isolation forest algorithm to detect anomalies in activity data, followed by the development of a recurrent neural network (RNN) model to predict these anomalies over time. The RNN model demonstrated high predictive accuracy, with a mean accuracy of 0.953 and a mean loss of 0.124 for participants with ADHD. These findings suggest that machine learning techniques, particularly RNNs, can effectively identify and predict anomalies in temporal motor activity data, offering objective insights into ADHD-related movement behaviors. This approach is promising for informing personalized interventions and improving clinical decision-making in the management of ADHD. |
Davide Ferraris Konstantinos Kotis, Christos Kalloniatis Enhancing TrUStAPIS Methodology in the Web of Things with LLM-generated IoT Trust Semantics Conference The 2024 International Conference on Information and Communications Security (ICICS 2024), 2024. @conference{Ferraris2024, title = {Enhancing TrUStAPIS Methodology in the Web of Things with LLM-generated IoT Trust Semantics}, author = {Davide Ferraris, Konstantinos Kotis, Christos Kalloniatis}, url = {https://link.springer.com/chapter/10.1007/978-981-97-8798-2_7}, doi = {https://doi.org/10.1007/978-981-97-8798-2_7}, year = {2024}, date = {2024-12-25}, booktitle = {The 2024 International Conference on Information and Communications Security (ICICS 2024)}, pages = {125-144}, abstract = {In the Internet of Things (IoT) there are ecosystems where their physical ’smart’ entities virtually interact with each other. Often, this interaction occurs among unknown entities, making trust an essential requirement to overcome uncertainty in several aspects of this interaction. However, trust is a complex concept, and incorporating it in IoT is still a challenging topic. For this reason, it is highly significant to specify and model trust in early stages of the System Development Life Cycle (SDLC) of IoT-integrated systems, thus enhancing the aforementioned task. TrUStAPIS is a requirements engineering methodology recently introduced for incorporating trust requirements during IoT-based system design. The scope of this paper is to provide an extension of TrUStAPIS by introducing IoT trust semantics compatible with the W3C Web of Things (WoT) recommendations generated with the assistance of Large Language Models (LLMs). Taking advantage of LLMs as a tool for integrating and refining existing methodologies, in this paper we present our work towards a revision of the TrUStAPIS methodology. In this work, we contribute a new conceptual model and a refined JSON-LD ontology that takes into account IoT trust semantics, providing eventually a valuable tool for software engineers to design and model IoT-based systems and services.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } In the Internet of Things (IoT) there are ecosystems where their physical ’smart’ entities virtually interact with each other. Often, this interaction occurs among unknown entities, making trust an essential requirement to overcome uncertainty in several aspects of this interaction. However, trust is a complex concept, and incorporating it in IoT is still a challenging topic. For this reason, it is highly significant to specify and model trust in early stages of the System Development Life Cycle (SDLC) of IoT-integrated systems, thus enhancing the aforementioned task. TrUStAPIS is a requirements engineering methodology recently introduced for incorporating trust requirements during IoT-based system design. The scope of this paper is to provide an extension of TrUStAPIS by introducing IoT trust semantics compatible with the W3C Web of Things (WoT) recommendations generated with the assistance of Large Language Models (LLMs). Taking advantage of LLMs as a tool for integrating and refining existing methodologies, in this paper we present our work towards a revision of the TrUStAPIS methodology. In this work, we contribute a new conceptual model and a refined JSON-LD ontology that takes into account IoT trust semantics, providing eventually a valuable tool for software engineers to design and model IoT-based systems and services. |
Andreas Soularidis Konstantinos Kotis, Myriam Lamolle Zakaria Mejdoul Gaëlle Lortal George Vouros LLM-Assisted Generation of SWRL Rules from Natural Language Conference 2024 International Conference on AI x Data and Knowledge Engineering (AIxDKE), 2024, ISBN: 979-8-3315-1704-5. @conference{Soularidis2024b, title = {LLM-Assisted Generation of SWRL Rules from Natural Language}, author = {Andreas Soularidis, Konstantinos Kotis, Myriam Lamolle, Zakaria Mejdoul, Gaëlle Lortal, George Vouros}, doi = {https://doi.org/10.1109/AIxDKE63520.2024.00008}, isbn = {979-8-3315-1704-5}, year = {2024}, date = {2024-12-11}, booktitle = {2024 International Conference on AI x Data and Knowledge Engineering (AIxDKE)}, abstract = {Recently, Large Language Models (LLMs) have attracted great attention due to their remarkable performance in human-like text generation and reasoning skills (although their memory and hallucination problems still remain key issues to tackle more efficiently). LLMs have been applied to various application domains, including Knowledge Graph (KG) generation, question and answering over KGs and text-to-SPARQL translation. In this work, we investigate the capabilities of LLMs in text-to-SWRL translation, i.e., translation of Natural Language (NL) rules into Semantic Web Rule Language (SWRL) rules, put in the context of an industrial Ontology Engineering (OE) environment called GLUON, presenting our first experimental results. The aim of this work is to identify the level of automation that is adequate for the LLM to generate well-formed SWRL rules, towards the development of an LLM-based framework, as a plugin to the GLUON OE environment. In this direction we leverage and combine the reasoning capabilities of GPT-4o model, the Retrieval-Augmented Generation (RAG) technology, and prompt engineering. We employ quantitative and qualitative metrics to evaluate the generated SWRL rules, focusing on the correct syntax and the level of human intervention.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Recently, Large Language Models (LLMs) have attracted great attention due to their remarkable performance in human-like text generation and reasoning skills (although their memory and hallucination problems still remain key issues to tackle more efficiently). LLMs have been applied to various application domains, including Knowledge Graph (KG) generation, question and answering over KGs and text-to-SPARQL translation. In this work, we investigate the capabilities of LLMs in text-to-SWRL translation, i.e., translation of Natural Language (NL) rules into Semantic Web Rule Language (SWRL) rules, put in the context of an industrial Ontology Engineering (OE) environment called GLUON, presenting our first experimental results. The aim of this work is to identify the level of automation that is adequate for the LLM to generate well-formed SWRL rules, towards the development of an LLM-based framework, as a plugin to the GLUON OE environment. In this direction we leverage and combine the reasoning capabilities of GPT-4o model, the Retrieval-Augmented Generation (RAG) technology, and prompt engineering. We employ quantitative and qualitative metrics to evaluate the generated SWRL rules, focusing on the correct syntax and the level of human intervention. |
Search with keywords
| 1. | Andreas Kontogiannis Vasilis Pollatos, Panayotis Mertikopoulos Ioannis Panageas : Efficient swap regret minimization in combinatorial bandits. Twenty-Ninth Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026), 2026. (Type: Conference | Abstract | Links | BibTeX) @conference{Kontogiannis2026, title = {Efficient swap regret minimization in combinatorial bandits}, author = {Andreas Kontogiannis, Vasilis Pollatos, Panayotis Mertikopoulos, Ioannis Panageas}, url = {https://arxiv.org/pdf/2602.02087}, year = {2026}, date = {2026-05-02}, booktitle = {Twenty-Ninth Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026)}, abstract = {This paper addresses the problem of designing efficient no-swap regret algorithms for combinatorial bandits, where the number of actions N is exponentially large in the dimensionality of the problem. In this setting, designing efficient no-swap regret translates to sublinear — in horizon T — swap regret with polylogarithmic dependence on N. In contrast to the weaker notion of external regret minimization – a problem which is fairly well understood in the literature – achieving no-swap regret with a polylogarithmic dependence on N has remained elusive in combinatorial bandits. Our paper resolves this challenge, by introducing a no-swap-regret learning algorithm with regret that scales polylogarithmically in N and is tight for the class of combinatorial bandits. To ground our results, we also demonstrate how to implement the proposed algorithm efficiently — that is, with a per-iteration complexity that also scales polylogarithmically in N — across a wide range of well-studied applications.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } This paper addresses the problem of designing efficient no-swap regret algorithms for combinatorial bandits, where the number of actions N is exponentially large in the dimensionality of the problem. In this setting, designing efficient no-swap regret translates to sublinear — in horizon T — swap regret with polylogarithmic dependence on N. In contrast to the weaker notion of external regret minimization – a problem which is fairly well understood in the literature – achieving no-swap regret with a polylogarithmic dependence on N has remained elusive in combinatorial bandits. Our paper resolves this challenge, by introducing a no-swap-regret learning algorithm with regret that scales polylogarithmically in N and is tight for the class of combinatorial bandits. To ground our results, we also demonstrate how to implement the proposed algorithm efficiently — that is, with a per-iteration complexity that also scales polylogarithmically in N — across a wide range of well-studied applications. |
| 2. | Georgios Bouchouras Dimitrios Doumanas, Andreas Soularidis Konstantinos Kotis George Vouros : Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting. In: AI, 7 (4), pp. 139, 2026, ISSN: 2673-2688. (Type: Journal Article | Abstract | Links | BibTeX) @article{Bouchouras2026, title = {Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting}, author = {Georgios Bouchouras, Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis, George Vouros}, url = {https://www.mdpi.com/2673-2688/7/4/139}, doi = {https://doi.org/10.3390/ai7040139}, issn = {2673-2688}, year = {2026}, date = {2026-04-14}, journal = {AI}, volume = {7}, number = {4}, pages = {139}, abstract = {Ontology engineering plays a critical role in clinical decision support systems for Parkinson’s Disease (PD) monitoring and alerting. While Large Language Models (LLMs) have shown promise in knowledge modeling tasks, their effectiveness in autonomously constructing comprehensive ontologies for complex clinical domains remains unclear. This study investigates four ontology engineering methodologies for PD monitoring and alerting: One-shot (OS) prompting, Decomposed Sequential Prompting (DSP), X-HCOME, and SimX-HCOME+. Multiple LLMs were evaluated across these methodologies. Generated ontologies were assessed against a reference PD ontology using structural evaluation metrics focused on classes and object properties. Expert review was additionally conducted to analyze knowledge extensions beyond the gold standard. LLMs were able to autonomously generate syntactically valid and semantically meaningful ontologies using OS and DSP prompting; however, these ontologies exhibited limited conceptual coverage. Incorporating human expertise through X-HCOME significantly improved ontology completeness and evaluation metrics. Expert review further validated clinically relevant concepts absent from the reference ontology. SimX-HCOME+ demonstrated that iterative, supervised collaboration supports ontology refinement, although challenges persisted in natural language-to-rule formalization. The findings suggest that LLMs are more effective as collaborative assistants rather than standalone ontology engineers in the PD domain. Structured human–LLM collaboration is associated with improved ontology coverage and facilitates the identification of potential knowledge extensions in clinical monitoring applications. While the present evaluation focuses primarily on structural ontology elements, the proposed methodologies provide useful insights for LLM-assisted ontology engineering in complex healthcare domains.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Ontology engineering plays a critical role in clinical decision support systems for Parkinson’s Disease (PD) monitoring and alerting. While Large Language Models (LLMs) have shown promise in knowledge modeling tasks, their effectiveness in autonomously constructing comprehensive ontologies for complex clinical domains remains unclear. This study investigates four ontology engineering methodologies for PD monitoring and alerting: One-shot (OS) prompting, Decomposed Sequential Prompting (DSP), X-HCOME, and SimX-HCOME+. Multiple LLMs were evaluated across these methodologies. Generated ontologies were assessed against a reference PD ontology using structural evaluation metrics focused on classes and object properties. Expert review was additionally conducted to analyze knowledge extensions beyond the gold standard. LLMs were able to autonomously generate syntactically valid and semantically meaningful ontologies using OS and DSP prompting; however, these ontologies exhibited limited conceptual coverage. Incorporating human expertise through X-HCOME significantly improved ontology completeness and evaluation metrics. Expert review further validated clinically relevant concepts absent from the reference ontology. SimX-HCOME+ demonstrated that iterative, supervised collaboration supports ontology refinement, although challenges persisted in natural language-to-rule formalization. The findings suggest that LLMs are more effective as collaborative assistants rather than standalone ontology engineers in the PD domain. Structured human–LLM collaboration is associated with improved ontology coverage and facilitates the identification of potential knowledge extensions in clinical monitoring applications. While the present evaluation focuses primarily on structural ontology elements, the proposed methodologies provide useful insights for LLM-assisted ontology engineering in complex healthcare domains. |
| 3. | Dimitrios Doumanas Andreas Soularidis, Nikolaos Zafeiropoulos Stamatis Chatzistamatis George Tsekouras Andreas El Saer Chrisaphis Nathanailidis Konstantinos Kotis E: Unbiasing Greek: In-Context Learning Strategies for Gender Bias Identification and Mitigation for Legal Documents and Job Ads. In: Information, 17 (4), pp. 342, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Doumanas2026b, title = {Unbiasing Greek: In-Context Learning Strategies for Gender Bias Identification and Mitigation for Legal Documents and Job Ads}, author = {Dimitrios Doumanas, Andreas Soularidis, Nikolaos Zafeiropoulos, Stamatis Chatzistamatis, George E Tsekouras, Andreas El Saer, Chrisaphis Nathanailidis, Konstantinos Kotis}, url = {https://www.mdpi.com/2078-2489/17/4/342}, doi = {https://doi.org/10.3390/info17040342}, year = {2026}, date = {2026-04-02}, journal = {Information}, volume = {17}, number = {4}, pages = {342}, abstract = {Gender bias embedded in legal and professional texts perpetuates systemic inequality, yet research on bias identification and mitigation remains largely confined to English. Morphologically rich languages such as Greek, where grammatical gender pervades nouns, adjectives, pronouns, and participles, present unique challenges that existing approaches fail to address. This paper elaborates on a systematic methodology primarily focusing on identifying and mitigating gender bias in Greek-language job advertisements and legal documents. To accomplish that task, we define a taxonomy of nine gender bias rules tailored to the linguistic properties of Greek and construct domain-specific annotated datasets comprising 90 expert-curated few-shot examples across both textual domains. Using these resources, we employ XML-structured prompt engineering with in-context learning (ICL)and systematically compare three classes of models: (i) commercial large language models (LLMs), namely Claude Sonnet 4.5 and GPT-5.2, (ii) two open-weight small language models (SLMs), Mistral Small (24B) and Ministral (14B), and (iii) Llama Krikri (8B), a Greek-native language model built on Llama 3.1 and fine-tuned on high-quality Greek corpora. For each input text, the system identifies biased expressions, maps them to specific bias rules, provides explanations, and generates a fully corrected inclusive version. Our experiments reveal substantial performance disparities across model scales and linguistic specialization, with LLMs demonstrating superior contextual reasoning and SLMs exhibiting systematic over-correction and grammatical errors in Greek morphology. We further introduce a critical meta-rule addressing gender agreement with named entities to prevent spurious corrections in legal texts referencing identified individuals. The findings highlight the importance of model scale, language-specific adaptation, and carefully designed prompting strategies for bias mitigation in underrepresented languages.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Gender bias embedded in legal and professional texts perpetuates systemic inequality, yet research on bias identification and mitigation remains largely confined to English. Morphologically rich languages such as Greek, where grammatical gender pervades nouns, adjectives, pronouns, and participles, present unique challenges that existing approaches fail to address. This paper elaborates on a systematic methodology primarily focusing on identifying and mitigating gender bias in Greek-language job advertisements and legal documents. To accomplish that task, we define a taxonomy of nine gender bias rules tailored to the linguistic properties of Greek and construct domain-specific annotated datasets comprising 90 expert-curated few-shot examples across both textual domains. Using these resources, we employ XML-structured prompt engineering with in-context learning (ICL)and systematically compare three classes of models: (i) commercial large language models (LLMs), namely Claude Sonnet 4.5 and GPT-5.2, (ii) two open-weight small language models (SLMs), Mistral Small (24B) and Ministral (14B), and (iii) Llama Krikri (8B), a Greek-native language model built on Llama 3.1 and fine-tuned on high-quality Greek corpora. For each input text, the system identifies biased expressions, maps them to specific bias rules, provides explanations, and generates a fully corrected inclusive version. Our experiments reveal substantial performance disparities across model scales and linguistic specialization, with LLMs demonstrating superior contextual reasoning and SLMs exhibiting systematic over-correction and grammatical errors in Greek morphology. We further introduce a critical meta-rule addressing gender agreement with named entities to prevent spurious corrections in legal texts referencing identified individuals. The findings highlight the importance of model scale, language-specific adaptation, and carefully designed prompting strategies for bias mitigation in underrepresented languages. |
| 4. | Andreas Kontogiannis Ioannis Panageas, Vasilis Pollatos : The computational complexity of avoiding strict saddle points in constrained optimization. In: arXiv, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Kontogiannis2026b, title = {The computational complexity of avoiding strict saddle points in constrained optimization}, author = {Andreas Kontogiannis, Ioannis Panageas, Vasilis Pollatos}, url = {https://arxiv.org/abs/2604.02285}, doi = {https://doi.org/10.48550/arXiv.2604.02285}, year = {2026}, date = {2026-04-02}, journal = {arXiv}, abstract = {While first-order stationary points (FOSPs) are the traditional targets of non-convex optimization, they often correspond to undesirable strict saddle points. To circumvent this, attention has shifted towards second-order stationary points (SOSPs). In unconstrained settings, finding approximate SOSPs is PLS-complete (Kontogiannis et al.), matching the complexity of finding unconstrained FOSPs (Hollender and Zampetakis). However, the complexity of finding SOSPs in constrained settings remained notoriously unclear and was highlighted as an important open question by both aforementioned works. Under one strict definition, even verifying whether a point is an approximate SOSP is NP-hard (Murty and Kabadi). Under another widely adopted, relaxed definition where non-negative curvature is required only along the null space of the active constraints, the problem lies in TFNP, and algorithms with O(poly(1/epsilon)) running times have been proposed (Lu et al.). In this work, we settle the complexity of constrained SOSP by proving that computing an epsilon-approximate SOSP under the tractable definition is PLS-complete. We demonstrate that our result holds even in the 2D unit square [0,1]^2, and remarkably, even when stationary points are isolated at a distance of Omega(1) from the domain’s boundary. Our result establishes a fundamental barrier: unless PLS is a subset of PPAD (implying PLS = CLS), no deterministic, iterative algorithm with an efficient, continuous update rule can exist for finding approximate SOSPs. This contrasts with the constrained first-order counterpart, for which Fearnley et al. showed that finding an approximate KKT point is CLS-complete. Finally, our result yields the first problem defined in a compact domain to be shown PLS-complete beyond the canonical Real-LocalOpt (Daskalakis and Papadimitriou).”}, keywords = {}, pubstate = {published}, tppubtype = {article} } While first-order stationary points (FOSPs) are the traditional targets of non-convex optimization, they often correspond to undesirable strict saddle points. To circumvent this, attention has shifted towards second-order stationary points (SOSPs). In unconstrained settings, finding approximate SOSPs is PLS-complete (Kontogiannis et al.), matching the complexity of finding unconstrained FOSPs (Hollender and Zampetakis). However, the complexity of finding SOSPs in constrained settings remained notoriously unclear and was highlighted as an important open question by both aforementioned works. Under one strict definition, even verifying whether a point is an approximate SOSP is NP-hard (Murty and Kabadi). Under another widely adopted, relaxed definition where non-negative curvature is required only along the null space of the active constraints, the problem lies in TFNP, and algorithms with O(poly(1/epsilon)) running times have been proposed (Lu et al.). In this work, we settle the complexity of constrained SOSP by proving that computing an epsilon-approximate SOSP under the tractable definition is PLS-complete. We demonstrate that our result holds even in the 2D unit square [0,1]^2, and remarkably, even when stationary points are isolated at a distance of Omega(1) from the domain’s boundary. Our result establishes a fundamental barrier: unless PLS is a subset of PPAD (implying PLS = CLS), no deterministic, iterative algorithm with an efficient, continuous update rule can exist for finding approximate SOSPs. This contrasts with the constrained first-order counterpart, for which Fearnley et al. showed that finding an approximate KKT point is CLS-complete. Finally, our result yields the first problem defined in a compact domain to be shown PLS-complete beyond the canonical Real-LocalOpt (Daskalakis and Papadimitriou)." |
| 5. | Georgios M Santipantakis Christos Doulkeridis, Petros Brimos : Semantic Data Transformation, FAIRification and Provenance for Data Spaces. In: Data in Brief, 66 , pp. 112675, 2026, ISSN: 2352-3409. (Type: Journal Article | Links | BibTeX) @article{Santipantakis2026, title = {Semantic Data Transformation, FAIRification and Provenance for Data Spaces}, author = {Georgios M Santipantakis, Christos Doulkeridis, Petros Brimos}, url = {https://www.sciencedirect.com/science/article/pii/S2352340926002283}, doi = {https://doi.org/10.1016/j.dib.2026.112675}, issn = {2352-3409}, year = {2026}, date = {2026-03-10}, journal = {Data in Brief}, volume = {66}, pages = {112675}, keywords = {}, pubstate = {published}, tppubtype = {article} } |
| 6. | Michael Kenteris, Konstantinos Kotis : The Convergence of Federated Learning, Knowledge Graphs, and Large Language Models for Language Learning: A Scoping Review. In: Applied Sciences, 16 (5), pp. 2611, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Kenteris2026, title = {The Convergence of Federated Learning, Knowledge Graphs, and Large Language Models for Language Learning: A Scoping Review}, author = {Michael Kenteris, Konstantinos Kotis}, url = {https://www.mdpi.com/2076-3417/16/5/2611}, doi = {https://doi.org/10.3390/app16052611}, year = {2026}, date = {2026-03-09}, journal = {Applied Sciences}, volume = {16}, number = {5}, pages = {2611}, abstract = {Large Language Models (LLMs) in Intelligent Computer-Assisted Language Learning enable highly personalized learning, yet raise significant challenges related to pedagogical grounding, data privacy, and instructional validity. Although Knowledge Graphs (KGs) and Federated Learning (FL) can mitigate these issues in isolation, evidence on systematic FL–KG–LLM integration for educational language learning remains limited. This scoping review maps the FL–KG–LLM convergence landscape. Following PRISMA-ScR guidelines, we searched six databases and screened 51 papers (2019–2025) using automated extraction. Our findings indicate limited convergence: no papers integrate all three domains, and 58.8% of approaches remain confined to isolated technological silos. Reporting is also uneven across the corpus, with an average “Not Reported” (NR) rate of 84.5%, most notably for privacy mechanisms (92.2%), validation metrics (90.2%), and Common European Framework of Reference for Languages (CEFR) alignment (88.2%). Domain-specific analysis reveals two distinct patterns: inter-domain gaps (disciplinary silos resulting in expected CEFR absence in single-domain papers) and intra-domain gaps (failure to report domain-critical variables, including 100% parameter NR in FL studies, 86.7% validation NR in KG studies, and 100% CEFR NR in convergence papers). Taken together, these gaps suggest that pedagogical grounding is treated as optional rather than structural. We therefore identify two pillars of pedagogical grounding: a Grounding Pillar, which constrains LLM outputs via Knowledge Graph rules, and a Validation Pillar, which concerns how authoritative frameworks (e.g., CEFR) are mapped onto Knowledge Graph schemas and evaluated. The near-universal absence of CEFR alignment and validation reporting suggests that this second pillar is currently missing, which we term the Integrity Gap—a systematic disconnection between technological innovation and pedagogical grounding inin Intelligent Computer-Assisted Language Learning. By reframing the problem as upstream control and validation, this review informs the design of user-facing automated systems where trust, transparency, and human oversight are critical.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Large Language Models (LLMs) in Intelligent Computer-Assisted Language Learning enable highly personalized learning, yet raise significant challenges related to pedagogical grounding, data privacy, and instructional validity. Although Knowledge Graphs (KGs) and Federated Learning (FL) can mitigate these issues in isolation, evidence on systematic FL–KG–LLM integration for educational language learning remains limited. This scoping review maps the FL–KG–LLM convergence landscape. Following PRISMA-ScR guidelines, we searched six databases and screened 51 papers (2019–2025) using automated extraction. Our findings indicate limited convergence: no papers integrate all three domains, and 58.8% of approaches remain confined to isolated technological silos. Reporting is also uneven across the corpus, with an average “Not Reported” (NR) rate of 84.5%, most notably for privacy mechanisms (92.2%), validation metrics (90.2%), and Common European Framework of Reference for Languages (CEFR) alignment (88.2%). Domain-specific analysis reveals two distinct patterns: inter-domain gaps (disciplinary silos resulting in expected CEFR absence in single-domain papers) and intra-domain gaps (failure to report domain-critical variables, including 100% parameter NR in FL studies, 86.7% validation NR in KG studies, and 100% CEFR NR in convergence papers). Taken together, these gaps suggest that pedagogical grounding is treated as optional rather than structural. We therefore identify two pillars of pedagogical grounding: a Grounding Pillar, which constrains LLM outputs via Knowledge Graph rules, and a Validation Pillar, which concerns how authoritative frameworks (e.g., CEFR) are mapped onto Knowledge Graph schemas and evaluated. The near-universal absence of CEFR alignment and validation reporting suggests that this second pillar is currently missing, which we term the Integrity Gap—a systematic disconnection between technological innovation and pedagogical grounding inin Intelligent Computer-Assisted Language Learning. By reframing the problem as upstream control and validation, this review informs the design of user-facing automated systems where trust, transparency, and human oversight are critical. |
| 7. | Elias Alevizos Georgios M Santipantakis, Christos Doulkeridis Alexander Artikis : Online spatial reasoning for complex event recognition. In: GeoInformatica, 30 (1), pp. 9, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Alevizos2026, title = {Online spatial reasoning for complex event recognition}, author = {Elias Alevizos, Georgios M Santipantakis, Christos Doulkeridis, Alexander Artikis}, url = {https://link.springer.com/article/10.1007/s10707-026-00569-z}, doi = {https://doi.org/10.1007/s10707-026-00569-z}, year = {2026}, date = {2026-03-03}, journal = {GeoInformatica}, volume = {30}, number = {1}, pages = {9}, abstract = {Complex Event Recognition (CER) systems have the ability to process streams of events by detecting event patterns with minimal latency. Typically, these patterns have a temporal structure, often resembling the sequential structure of regular expressions. A pattern advances to the next state by checking various conditions on the current and possibly previous events of the stream. CER systems are very efficient in tracking all the possible paths that a pattern may follow and report when a path is complete and a complex event must be reported. In some cases, the conditions that need to be checked may be spatial. For example, in maritime situational awareness, a condition may need to check whether a vessel is close to any other vessel. Such conditions are not easily expressed directly as regular expressions. For such spatio-temporal tasks, there exist dedicated modules which can evaluate this type of conditions efficiently. Thus, we can integrate such a spatio-temporal module within a CER system in order to take advantage of both worlds: the CER engine can accommodate and process complex regular expressions and delegate the evaluation of expensive spatio-temporal tasks to a dedicated module whenever it needs to. We present an approach towards such an integration. We describe how a CER engine, based on symbolic automata, can cooperate with a spatio-temporal link discovery (stLD) module such that the former can leverage the spatio-temporal capabilities of the latter. This cooperation can take place in an online manner rendering the whole system suitable for real-time processing of event streams. We discuss two different communication schemes between the CER engine and the spatio-temporal module and explore when each one should be preferred. We provide a theoretical estimation of the predicted performance of the system under each communication scheme. Our extensive experimental evaluation confirms most of our theoretical predictions.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Complex Event Recognition (CER) systems have the ability to process streams of events by detecting event patterns with minimal latency. Typically, these patterns have a temporal structure, often resembling the sequential structure of regular expressions. A pattern advances to the next state by checking various conditions on the current and possibly previous events of the stream. CER systems are very efficient in tracking all the possible paths that a pattern may follow and report when a path is complete and a complex event must be reported. In some cases, the conditions that need to be checked may be spatial. For example, in maritime situational awareness, a condition may need to check whether a vessel is close to any other vessel. Such conditions are not easily expressed directly as regular expressions. For such spatio-temporal tasks, there exist dedicated modules which can evaluate this type of conditions efficiently. Thus, we can integrate such a spatio-temporal module within a CER system in order to take advantage of both worlds: the CER engine can accommodate and process complex regular expressions and delegate the evaluation of expensive spatio-temporal tasks to a dedicated module whenever it needs to. We present an approach towards such an integration. We describe how a CER engine, based on symbolic automata, can cooperate with a spatio-temporal link discovery (stLD) module such that the former can leverage the spatio-temporal capabilities of the latter. This cooperation can take place in an online manner rendering the whole system suitable for real-time processing of event streams. We discuss two different communication schemes between the CER engine and the spatio-temporal module and explore when each one should be preferred. We provide a theoretical estimation of the predicted performance of the system under each communication scheme. Our extensive experimental evaluation confirms most of our theoretical predictions. |
| 8. | George Papadopoulos, George Vouros A: Learning to maintain safety through expert demonstrations in settings with unknown constraints: A Q-learning perspective. In: arXiv, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Papadopoulos2026, title = {Learning to maintain safety through expert demonstrations in settings with unknown constraints: A Q-learning perspective}, author = {George Papadopoulos, George A Vouros}, url = {https://arxiv.org/abs/2602.23816 https://arxiv.org/pdf/2602.23816}, doi = {https://doi.org/10.48550/arXiv.2602.23816}, year = {2026}, date = {2026-02-27}, journal = {arXiv}, abstract = {Given a set of trajectories demonstrating the execution of a task safely in a constrained MDP with observable rewards but with unknown constraints and non-observable costs, we aim to find a policy that maximizes the likelihood of demonstrated trajectories trading the balance between being conservative and increasing significantly the likelihood of high-rewarding trajectories but with potentially unsafe steps. Having these objectives, we aim towards learning a policy that maximizes the probability of the most $promising$ trajectories with respect to the demonstrations. In so doing, we formulate the “promise” of individual state-action pairs in terms of $Q$ values, which depend on task-specific rewards as well as on the assessment of states’ safety, mixing expectations in terms of rewards and safety. This entails a safe Q-learning perspective of the inverse learning problem under constraints: The devised Safe $Q$ Inverse Constrained Reinforcement Learning (SafeQIL) algorithm is compared to state-of-the art inverse constraint reinforcement learning algorithms to a set of challenging benchmark tasks, showing its merits.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Given a set of trajectories demonstrating the execution of a task safely in a constrained MDP with observable rewards but with unknown constraints and non-observable costs, we aim to find a policy that maximizes the likelihood of demonstrated trajectories trading the balance between being conservative and increasing significantly the likelihood of high-rewarding trajectories but with potentially unsafe steps. Having these objectives, we aim towards learning a policy that maximizes the probability of the most $promising$ trajectories with respect to the demonstrations. In so doing, we formulate the “promise" of individual state-action pairs in terms of $Q$ values, which depend on task-specific rewards as well as on the assessment of states’ safety, mixing expectations in terms of rewards and safety. This entails a safe Q-learning perspective of the inverse learning problem under constraints: The devised Safe $Q$ Inverse Constrained Reinforcement Learning (SafeQIL) algorithm is compared to state-of-the art inverse constraint reinforcement learning algorithms to a set of challenging benchmark tasks, showing its merits. |
| 9. | Dimitrios Doumanas, Konstantinos Kotis : ReaDS-KG: An LLM-Knowledge Graph Framework for Reasoned Decision Support in Dynamic Safety-Critical Domains. In: TechRxiv, 2026. (Type: Journal Article | Abstract | Links | BibTeX) @article{Doumanas2026, title = {ReaDS-KG: An LLM-Knowledge Graph Framework for Reasoned Decision Support in Dynamic Safety-Critical Domains}, author = {Dimitrios Doumanas, Konstantinos Kotis}, url = {https://www.techrxiv.org/doi/full/10.36227/techrxiv.176826793.34811491/v1}, doi = {https://doi.org/10.36227/techrxiv.176826793.34811491/v1}, year = {2026}, date = {2026-01-13}, journal = {TechRxiv}, abstract = {Safety-critical domains such as military operations, border security, and search-and-rescue must operate under uncertainty, severe time pressure, and continuously changing conditions. In these settings, decision-support systems must not only provide accurate recommendations but also make the underlying reasoning explicit and auditable. This paper introduces ReaDS-KG (Reasoned Decision Support over Knowledge Graphs), an LLM-Knowledge Graph framework that delivers reasoned rather than purely predictive support. ReaDS-KG represents domain knowledge, assets, constraints, and causal dependencies in an ontology-driven knowledge graph, and uses a large language model to (i) translate natural-language questions into Cypher queries, (ii) orchestrate graph-based reasoning over causal structures, and (iii) return narrative answers with explicit justifications grounded in the graph. The framework follows a five-stage pipeline: ontology design, data-to-KG transformation, causal enrichment, LLM-mediated querying, and scenariobased evaluation. To demonstrate its applicability, we instantiate ReaDS-KG in a synthetic brigade-level operational scenario and pose twenty decision-oriented questions, covering feasibility, mobility, sustainment, command-and-control robustness, and risk. We then compare an LLM+KG agent powered by ReaDS-KG to ten active-duty officers using an eight-dimensional scoring rubric. The agent achieves decision-support quality comparable to field-grade officers and clearly above junior officers, while responding at machine response speed and providing transparent reasoning chains. These results suggest that ReaDS-KG can function as a quasi-expert, explainable staff assistant in dynamic safety-critical domains, and the architecture is readily transferable to other safety-critical settings that share similar uncertainty and causal-reasoning requirements, such as border management and disaster response.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Safety-critical domains such as military operations, border security, and search-and-rescue must operate under uncertainty, severe time pressure, and continuously changing conditions. In these settings, decision-support systems must not only provide accurate recommendations but also make the underlying reasoning explicit and auditable. This paper introduces ReaDS-KG (Reasoned Decision Support over Knowledge Graphs), an LLM-Knowledge Graph framework that delivers reasoned rather than purely predictive support. ReaDS-KG represents domain knowledge, assets, constraints, and causal dependencies in an ontology-driven knowledge graph, and uses a large language model to (i) translate natural-language questions into Cypher queries, (ii) orchestrate graph-based reasoning over causal structures, and (iii) return narrative answers with explicit justifications grounded in the graph. The framework follows a five-stage pipeline: ontology design, data-to-KG transformation, causal enrichment, LLM-mediated querying, and scenariobased evaluation. To demonstrate its applicability, we instantiate ReaDS-KG in a synthetic brigade-level operational scenario and pose twenty decision-oriented questions, covering feasibility, mobility, sustainment, command-and-control robustness, and risk. We then compare an LLM+KG agent powered by ReaDS-KG to ten active-duty officers using an eight-dimensional scoring rubric. The agent achieves decision-support quality comparable to field-grade officers and clearly above junior officers, while responding at machine response speed and providing transparent reasoning chains. These results suggest that ReaDS-KG can function as a quasi-expert, explainable staff assistant in dynamic safety-critical domains, and the architecture is readily transferable to other safety-critical settings that share similar uncertainty and causal-reasoning requirements, such as border management and disaster response. |
| 10. | Andreas Soularidis Dimitrios Doumanas, Konstantinos Kotis George Vouros A: Automating agentic collaborative ontology engineering with role-playing simulation of LLM-powered agents and RAG technology. In: The Knowledge Engineering Review, 40 , pp. e10, 2025. (Type: Journal Article | Abstract | Links | BibTeX) @article{Soularidis2025, title = {Automating agentic collaborative ontology engineering with role-playing simulation of LLM-powered agents and RAG technology}, author = {Andreas Soularidis, Dimitrios Doumanas, Konstantinos Kotis, George A Vouros}, doi = {https://doi.org/10.1017/S026988892510009X}, year = {2025}, date = {2025-12-19}, journal = {The Knowledge Engineering Review}, volume = {40}, pages = {e10}, abstract = {Motivated by the astonishing capabilities of large language models (LLMs) in text-generation, reasoning, and simulation of complex human behaviors, in this paper, we propose a novel multi-component LLM-based framework, namely LLM4ACOE, that fully automates the collaborative ontology engineering (COE) process using role-playing simulation of LLM agents and retrieval augmented generation (RAG) technology. The proposed solution enhances the LLM-powered role-playing simulation with RAG ‘feeding’ the LLM with three different types of external knowledge. This knowledge corresponds to the knowledge required by each of the COE roles (agents), using a component-based framework, as follows: (a) domain-specific data-centric documents, (b) OWL documentation, and (c) ReAct guidelines. The aforementioned components are evaluated in combination, with the aim of investigating their impact on the quality of generated ontologies. The aim of this work is twofold, (a) to identify the capacity of LLM-based agents to generate acceptable (by human-experts) ontologies through agentic collaborative ontology engineering (ACOE) role-playing simulation, at specific levels of acceptance (accuracy, validity, and expressiveness of ontologies) without human intervention and (b) to investigate whether and/or to what extent the selected RAG components affect the quality of the generated ontologies. The evaluation of this novel approach is performed using ChatGPT-o in the domain of search and rescue (SAR) missions. To assess the generated ontologies, quantitative and qualitative measures are employed, focusing on coverage, expressiveness, structure, and human involvement.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Motivated by the astonishing capabilities of large language models (LLMs) in text-generation, reasoning, and simulation of complex human behaviors, in this paper, we propose a novel multi-component LLM-based framework, namely LLM4ACOE, that fully automates the collaborative ontology engineering (COE) process using role-playing simulation of LLM agents and retrieval augmented generation (RAG) technology. The proposed solution enhances the LLM-powered role-playing simulation with RAG ‘feeding’ the LLM with three different types of external knowledge. This knowledge corresponds to the knowledge required by each of the COE roles (agents), using a component-based framework, as follows: (a) domain-specific data-centric documents, (b) OWL documentation, and (c) ReAct guidelines. The aforementioned components are evaluated in combination, with the aim of investigating their impact on the quality of generated ontologies. The aim of this work is twofold, (a) to identify the capacity of LLM-based agents to generate acceptable (by human-experts) ontologies through agentic collaborative ontology engineering (ACOE) role-playing simulation, at specific levels of acceptance (accuracy, validity, and expressiveness of ontologies) without human intervention and (b) to investigate whether and/or to what extent the selected RAG components affect the quality of the generated ontologies. The evaluation of this novel approach is performed using ChatGPT-o in the domain of search and rescue (SAR) missions. To assess the generated ontologies, quantitative and qualitative measures are employed, focusing on coverage, expressiveness, structure, and human involvement. |
| 11. | Apostolos Glenis, George Vouros : Scalable Univariate and Multivariate Time-Series Classifiers with Deep Learning Methods Exploiting Symbolic Representations. In: Computers, 14 (12), pp. 563, 2025. (Type: Journal Article | Abstract | Links | BibTeX) @article{Glenis2025, title = {Scalable Univariate and Multivariate Time-Series Classifiers with Deep Learning Methods Exploiting Symbolic Representations}, author = {Apostolos Glenis, George Vouros}, url = {https://www.mdpi.com/2073-431X/14/12/563}, doi = {https://doi.org/10.3390/computers14120563}, year = {2025}, date = {2025-12-17}, journal = {Computers}, volume = {14}, number = {12}, pages = {563}, abstract = {Time-series classification (TSC) is an important task across sciences. Symbolic representations (especially SFA) are very effective at combating noise. In this paper, we employ symbolic representations to create state-of-the-art time-series classifiers, with the aim to advance scalability without sacrificing accuracy. First, we create a graph representation of the time series based on SFA words. We use this representation together with graph kernels and an SVM classifier to create a scalable time-series classifier. Next, we use the graph representation together with a Graph Convolutional Neural Network to test how it fares against state-of-the-art time-series classifiers. Additionally, we devised deep neural networks exploiting the SFA representation, inspired by the text classification domain, to study how they fare against state-of-the-art classifiers. The proposed deep learning classifiers have been adapted and evaluated for the multivariate time-series case and also against state-of-the-art time-series classification algorithms based on symbolic representations.}, keywords = {}, pubstate = {published}, tppubtype = {article} } Time-series classification (TSC) is an important task across sciences. Symbolic representations (especially SFA) are very effective at combating noise. In this paper, we employ symbolic representations to create state-of-the-art time-series classifiers, with the aim to advance scalability without sacrificing accuracy. First, we create a graph representation of the time series based on SFA words. We use this representation together with graph kernels and an SVM classifier to create a scalable time-series classifier. Next, we use the graph representation together with a Graph Convolutional Neural Network to test how it fares against state-of-the-art time-series classifiers. Additionally, we devised deep neural networks exploiting the SFA representation, inspired by the text classification domain, to study how they fare against state-of-the-art classifiers. The proposed deep learning classifiers have been adapted and evaluated for the multivariate time-series case and also against state-of-the-art time-series classification algorithms based on symbolic representations. |
| 12. | Georgios Bouchouras Dimitrios Doumanas, Andreas Soularidis Konstantinos Kotis George Vouros A: Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting. In: arXiv, 2025. (Type: Journal Article | Abstract | Links | BibTeX) @article{Bouchouras2025, title = {Leveraging LLMs for Collaborative Ontology Engineering in Parkinson Disease Monitoring and Alerting}, author = {Georgios Bouchouras, Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis, George A Vouros}, url = {https://arxiv.org/pdf/2512.14288}, doi = {https://doi.org/10.48550/arXiv.2512.14288}, year = {2025}, date = {2025-12-16}, journal = {arXiv}, abstract = {This paper explores the integration of Large Language Models (LLMs) in the engineering of a Parkinson’s Disease (PD) monitoring and alerting ontology through four key methodologies: One Shot (OS) prompt techniques, Chain of Thought (CoT) prompts, X-HCOME, and SimX-HCOME+. The primary objective is to determine whether LLMs alone can create comprehensive ontologies and, if not, whether human-LLM collaboration can achieve this goal. Consequently, the paper assesses the effectiveness of LLMs in automated ontology development and the enhancement achieved through human-LLM collaboration. Initial ontology generation was performed using One Shot (OS) and Chain of Thought (CoT) prompts, demonstrating the capability of LLMs to autonomously construct ontologies for PD monitoring and alerting. However, these outputs were not comprehensive and required substantial human refinement to enhance their completeness and accuracy. X-HCOME, a hybrid ontology engineering approach that combines human expertise with LLM capabilities, showed significant improvements in ontology comprehensiveness. This methodology resulted in ontologies that are very similar to those constructed by experts. Further experimentation with SimX-HCOME+, another hybrid methodology emphasizing continuous human supervision and iterative refinement, highlighted the importance of ongoing human involvement. This approach led to the creation of more comprehensive and accurate ontologies. Overall, the paper underscores the potential of human-LLM collaboration in advancing ontology engineering, particularly in complex domains like PD. The results suggest promising directions for future research, including the development of specialized GPT models for ontology construction.}, keywords = {}, pubstate = {published}, tppubtype = {article} } This paper explores the integration of Large Language Models (LLMs) in the engineering of a Parkinson’s Disease (PD) monitoring and alerting ontology through four key methodologies: One Shot (OS) prompt techniques, Chain of Thought (CoT) prompts, X-HCOME, and SimX-HCOME+. The primary objective is to determine whether LLMs alone can create comprehensive ontologies and, if not, whether human-LLM collaboration can achieve this goal. Consequently, the paper assesses the effectiveness of LLMs in automated ontology development and the enhancement achieved through human-LLM collaboration. Initial ontology generation was performed using One Shot (OS) and Chain of Thought (CoT) prompts, demonstrating the capability of LLMs to autonomously construct ontologies for PD monitoring and alerting. However, these outputs were not comprehensive and required substantial human refinement to enhance their completeness and accuracy. X-HCOME, a hybrid ontology engineering approach that combines human expertise with LLM capabilities, showed significant improvements in ontology comprehensiveness. This methodology resulted in ontologies that are very similar to those constructed by experts. Further experimentation with SimX-HCOME+, another hybrid methodology emphasizing continuous human supervision and iterative refinement, highlighted the importance of ongoing human involvement. This approach led to the creation of more comprehensive and accurate ontologies. Overall, the paper underscores the potential of human-LLM collaboration in advancing ontology engineering, particularly in complex domains like PD. The results suggest promising directions for future research, including the development of specialized GPT models for ontology construction. |
| 13. | Asimina Dimara Konstantinos Kotis, Stamatis Chatzistamatis Nikolaos Evangeliou Chrysaphis Nathanailidis George Tsekouras E: Towards Effective Data Process Pipelines for Legal NLP in English and Non-English Languages: A Greek Case Study. Computing, Communications and IoT Applications (ComComAp), 2025, ISBN: 979-8-3315-9143-4. (Type: Conference | Abstract | Links | BibTeX) @conference{Dimara2025b, title = {Towards Effective Data Process Pipelines for Legal NLP in English and Non-English Languages: A Greek Case Study}, author = {Asimina Dimara, Konstantinos Kotis, Stamatis Chatzistamatis, Nikolaos Evangeliou, Chrysaphis Nathanailidis, George E Tsekouras}, url = {https://ieeexplore.ieee.org/abstract/document/11353184}, doi = {https://doi.org/10.1109/ComComAp68359.2025.11353184}, isbn = {979-8-3315-9143-4}, year = {2025}, date = {2025-12-14}, booktitle = {Computing, Communications and IoT Applications (ComComAp)}, abstract = {Natural Language Processing (NLP) pipelines form the backbone of legal artificial intelligence applications, yet most existing tools are designed for English corpora and perform poorly when transferred to morphologically rich, non-English languages. This paper investigates these limitations through a comparative study of English and Greek legal texts. It is shown that English-centric pipelines exhibit systematic errors in preprocessing (tokenization, lemmatization, stop-word removal) and fail to capture legal semantics in embeddings, resulting in degraded downstream performance. To address these issues, a generalized framework is proposed that introduces language-specific preprocessing, curated legal resources, and multilingual embeddings fine-tuned on legal corpora. A case study demonstrates how adapted tools substantially improve similarity scores and classification accuracy in Greek legal texts, while highlighting persistent challenges such as grammatical gender bias. The findings underscore the need for fairness-aware, language-specific NLP pipelines to support robust and inclusive legal AI across diverse jurisdictions.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Natural Language Processing (NLP) pipelines form the backbone of legal artificial intelligence applications, yet most existing tools are designed for English corpora and perform poorly when transferred to morphologically rich, non-English languages. This paper investigates these limitations through a comparative study of English and Greek legal texts. It is shown that English-centric pipelines exhibit systematic errors in preprocessing (tokenization, lemmatization, stop-word removal) and fail to capture legal semantics in embeddings, resulting in degraded downstream performance. To address these issues, a generalized framework is proposed that introduces language-specific preprocessing, curated legal resources, and multilingual embeddings fine-tuned on legal corpora. A case study demonstrates how adapted tools substantially improve similarity scores and classification accuracy in Greek legal texts, while highlighting persistent challenges such as grammatical gender bias. The findings underscore the need for fairness-aware, language-specific NLP pipelines to support robust and inclusive legal AI across diverse jurisdictions. |
| 14. | Alexandros Karakikes, Konstantinos Kotis : AI-Assisted OSINT/SOCMINT for Safeguarding Borders: A Systematic Review. In: Information, 16 (12), pp. 1095, 2025, ISSN: 2078-2489. (Type: Journal Article | Abstract | Links | BibTeX) @article{Karakikes2025, title = {AI-Assisted OSINT/SOCMINT for Safeguarding Borders: A Systematic Review}, author = {Alexandros Karakikes, Konstantinos Kotis}, url = {https://www.mdpi.com/2078-2489/16/12/1095}, doi = {https://doi.org/10.3390/info16121095}, issn = {2078-2489}, year = {2025}, date = {2025-12-10}, journal = {Information}, volume = {16}, number = {12}, pages = {1095}, abstract = {In the highly volatile realm of global security, the necessity for leading-edge and effectual border resilience tactics has never been more imperative. This PRISMA 2020 guided systematic literature review (SLR) examines the intersection of artificial intelligence (AI), open-source intelligence (OSINT), and social media intelligence (SOCMINT) for enhancing border protection. Our systematic investigation across major databases (IEEE Xplore, Scopus, SpringerLink, MDPI, ACM) and grey literature sources yielded 3932 initial records and, after screening and eligibility assessment, 73 studies and reports from acknowledged organizations, contributing to the evidence synthesis. Three research questions (RQ1–RQ3) were addressed concerning the following: (a) the effectiveness and application of AI in OSINT/SOCMINT for border protection, its (b) data, technical, and operational limitations, and its (c) ethical, legal, and societal implications (GELSI). Evidence matrices summarize the findings, while narrative syntheses underline and thematically group the extracted insights. Results indicate that AI techniques—fluctuating from machine learning (ML) and natural language processing (NLP) to computer vision and emerging large language models (LLMs)—produce quantifiable improvements in forecasting irregular migration, detecting human trafficking, and supporting multimodal intelligence fusion. However, limitations include misinformation, data bias, adversarial vulnerabilities, governance deficits, and sandbox-to-production gaps. Ethical and societal concerns highlight risks of surveillance overreach, discrimination, and insufficient oversight, among others. To our knowledge, this is the first SLR at this intersection. We conclude that, AI-assisted OSINT/SOCMINT presents transformative potential for border protection requiring, nonetheless, balanced governance, robust validation, and future research on LLM/agentic AI, human–AI teaming, and oversight mechanisms.}, keywords = {}, pubstate = {published}, tppubtype = {article} } In the highly volatile realm of global security, the necessity for leading-edge and effectual border resilience tactics has never been more imperative. This PRISMA 2020 guided systematic literature review (SLR) examines the intersection of artificial intelligence (AI), open-source intelligence (OSINT), and social media intelligence (SOCMINT) for enhancing border protection. Our systematic investigation across major databases (IEEE Xplore, Scopus, SpringerLink, MDPI, ACM) and grey literature sources yielded 3932 initial records and, after screening and eligibility assessment, 73 studies and reports from acknowledged organizations, contributing to the evidence synthesis. Three research questions (RQ1–RQ3) were addressed concerning the following: (a) the effectiveness and application of AI in OSINT/SOCMINT for border protection, its (b) data, technical, and operational limitations, and its (c) ethical, legal, and societal implications (GELSI). Evidence matrices summarize the findings, while narrative syntheses underline and thematically group the extracted insights. Results indicate that AI techniques—fluctuating from machine learning (ML) and natural language processing (NLP) to computer vision and emerging large language models (LLMs)—produce quantifiable improvements in forecasting irregular migration, detecting human trafficking, and supporting multimodal intelligence fusion. However, limitations include misinformation, data bias, adversarial vulnerabilities, governance deficits, and sandbox-to-production gaps. Ethical and societal concerns highlight risks of surveillance overreach, discrimination, and insufficient oversight, among others. To our knowledge, this is the first SLR at this intersection. We conclude that, AI-assisted OSINT/SOCMINT presents transformative potential for border protection requiring, nonetheless, balanced governance, robust validation, and future research on LLM/agentic AI, human–AI teaming, and oversight mechanisms. |
| 15. | Dimitris Kostadimas Vlasios Kasapakis, Konstantinos Kotis : Exploiting VR, AIoT and Semantics Towards an Adaptive Virtual Museum. 20th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP), 2025, ISBN: 979-8-3315-8704-8. (Type: Conference | Abstract | Links | BibTeX) @conference{Kostadimas2025b, title = {Exploiting VR, AIoT and Semantics Towards an Adaptive Virtual Museum}, author = {Dimitris Kostadimas, Vlasios Kasapakis, Konstantinos Kotis}, url = {https://ieeexplore.ieee.org/abstract/document/11309793}, doi = {https://doi.org/10.1109/SMAP66932.2025.00034}, isbn = {979-8-3315-8704-8}, year = {2025}, date = {2025-11-27}, booktitle = {20th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP)}, pages = {157-162}, abstract = {Museums have long been spaces of wonder and discovery, but as technology evolves, so do the ways we engage with these cultural treasures. The design of adaptive virtual environments becomes essential to maintaining user interest and relevance. In this paper, an adaptive virtual museum system is proposed that explores the use of virtual reality (VR), artificial intelligence (AI), Internet of Things (IoT) as well as semantics to personalize and optimize virtual exhibition experiences. Based on the results of our previous research conducted regarding the possible combination of VR, AI and IoT (AIoT) for the design of innovative intelligent systems in different domains, our current work proposes a novel way to integrate all these technologies within the domain of cultural heritage (CH), a combination that remains relatively underexplored. The proposed framework, which is currently a work in progress, introduces new ways to modeling museums’ visitor behavior and preferences (mainly by using head-mounted displays (HMDs)) in a VR environment to dynamically adapt exhibition layouts, as well as to provide personalized content through a digital twin (DT) of a real museum. A key focus lies in intelligent user profiling and route/layout optimization to enhance visitor engagement and provide rich content through integration of Large Language Models (LLM). Although implementation is ongoing, this paper describes the conceptual design, core objectives, and anticipated impact on the broader scope of adaptive multimedia applications and personalized cultural experiences.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Museums have long been spaces of wonder and discovery, but as technology evolves, so do the ways we engage with these cultural treasures. The design of adaptive virtual environments becomes essential to maintaining user interest and relevance. In this paper, an adaptive virtual museum system is proposed that explores the use of virtual reality (VR), artificial intelligence (AI), Internet of Things (IoT) as well as semantics to personalize and optimize virtual exhibition experiences. Based on the results of our previous research conducted regarding the possible combination of VR, AI and IoT (AIoT) for the design of innovative intelligent systems in different domains, our current work proposes a novel way to integrate all these technologies within the domain of cultural heritage (CH), a combination that remains relatively underexplored. The proposed framework, which is currently a work in progress, introduces new ways to modeling museums’ visitor behavior and preferences (mainly by using head-mounted displays (HMDs)) in a VR environment to dynamically adapt exhibition layouts, as well as to provide personalized content through a digital twin (DT) of a real museum. A key focus lies in intelligent user profiling and route/layout optimization to enhance visitor engagement and provide rich content through integration of Large Language Models (LLM). Although implementation is ongoing, this paper describes the conceptual design, core objectives, and anticipated impact on the broader scope of adaptive multimedia applications and personalized cultural experiences. |
| 16. | Andreas Sideras Konstantinos Bougiatiotis, Elias Zavitsanos Georgios Paliouras George Vouros : A Multimodal Alignment-Based Anomaly Detection Method for Bankruptcy Prediction. Proceedings of the 6th ACM International Conference on AI in Finance, 2025, ISBN: 9798400722202. (Type: Conference | Abstract | Links | BibTeX) @conference{Sideras2025, title = {A Multimodal Alignment-Based Anomaly Detection Method for Bankruptcy Prediction}, author = {Andreas Sideras, Konstantinos Bougiatiotis, Elias Zavitsanos, Georgios Paliouras, George Vouros}, url = {https://dl.acm.org/doi/full/10.1145/3768292.3770380}, doi = {https://doi.org/10.1145/3768292.3770380}, isbn = {9798400722202}, year = {2025}, date = {2025-11-15}, booktitle = {Proceedings of the 6th ACM International Conference on AI in Finance}, pages = {53-61}, abstract = {We present a novel anomaly detection method for next-year bankruptcy prediction, utilizing a combination of financial figures and textual content from annual reports. Our approach, MABAD, learns a shared representation space where non-bankrupt firms share position and orientation. Samples that deviate from this pattern are assigned a higher anomaly score. The proposed method is tailored for highly imbalanced scenarios and is robust to heterogeneous, incomplete, and potentially contradictory inputs. We demonstrate that MABAD consistently outperforms a range of strong baselines, and we also curate and release a new publicly available multisource dataset to foster further research in the domain.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } We present a novel anomaly detection method for next-year bankruptcy prediction, utilizing a combination of financial figures and textual content from annual reports. Our approach, MABAD, learns a shared representation space where non-bankrupt firms share position and orientation. Samples that deviate from this pattern are assigned a higher anomaly score. The proposed method is tailored for highly imbalanced scenarios and is robust to heterogeneous, incomplete, and potentially contradictory inputs. We demonstrate that MABAD consistently outperforms a range of strong baselines, and we also curate and release a new publicly available multisource dataset to foster further research in the domain. |
| 17. | Elias Zavitsanos Konstantinos Bougiatiotis, Andreas Sideras Georgios Paliouras : Positive-Unlabeled Learning for Financial Misstatement Detection under Realistic Constraints. ICAIF ’25: Proceedings of the 6th ACM International Conference on AI in Finance, 2025, ISBN: 9798400722202. (Type: Conference | Abstract | Links | BibTeX) @conference{Zavitsanos2025, title = {Positive-Unlabeled Learning for Financial Misstatement Detection under Realistic Constraints}, author = {Elias Zavitsanos, Konstantinos Bougiatiotis, Andreas Sideras, Georgios Paliouras}, url = {https://dl.acm.org/doi/full/10.1145/3768292.3770366 https://dl.acm.org/doi/epdf/10.1145/3768292.3770366}, doi = {https://doi.org/10.1145/3768292.3770366}, isbn = {9798400722202}, year = {2025}, date = {2025-11-15}, booktitle = {ICAIF ’25: Proceedings of the 6th ACM International Conference on AI in Finance}, pages = {864-872}, abstract = {Detecting financial misstatements is critical for market integrity but remains challenging due to class imbalance, delayed discovery, and limited labeled data. We propose a novel Positive-Unlabeled (PU) learning framework that models the detection task under realistic constraints, where only a small subset of misstatements is known at training time. Our approach integrates unlabeled data into training, preserves temporal structure, and accounts for extreme imbalance. We construct and release a benchmark dataset reflecting these characteristics and evaluate several PU learning methods against recent baselines. Results show that PU-based models consistently outperform supervised approaches, highlighting their suitability for real-world misstatement detection.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Detecting financial misstatements is critical for market integrity but remains challenging due to class imbalance, delayed discovery, and limited labeled data. We propose a novel Positive-Unlabeled (PU) learning framework that models the detection task under realistic constraints, where only a small subset of misstatements is known at training time. Our approach integrates unlabeled data into training, preserves temporal structure, and accounts for extreme imbalance. We construct and release a benchmark dataset reflecting these characteristics and evaluate several PU learning methods against recent baselines. Results show that PU-based models consistently outperform supervised approaches, highlighting their suitability for real-world misstatement detection. |
| 18. | Dimitrios Doumanas Andreas Soularidis, Konstantinos Kotis : Causal Reasoning and Large Language Models for Military Decision-Making: Rethinking the Command Structures in the Era of Generative AI. In: AI, 7 (1), pp. 14, 2025, ISSN: 2673-2688. (Type: Journal Article | Abstract | Links | BibTeX) @article{Doumanas2025e, title = {Causal Reasoning and Large Language Models for Military Decision-Making: Rethinking the Command Structures in the Era of Generative AI}, author = {Dimitrios Doumanas, Andreas Soularidis, Konstantinos Kotis}, url = {https://www.mdpi.com/2673-2688/7/1/14}, doi = {https://doi.org/10.3390/ai7010014}, issn = {2673-2688}, year = {2025}, date = {2025-10-24}, journal = {AI}, volume = {7}, number = {1}, pages = {14}, abstract = {Military decision-making is inherently complex and highly critical, requiring commanders to assess multiple variables in real-time, anticipate second-order effects, and adapt strategies based on continuously evolving battlefield conditions. Traditional approaches rely on domain expertise, experience, and intuition, often supported by decision-support systems designed by military experts. With the rapid advancement of Large Language Models (LLMs) such as ChatGPT, Claude, and DeepSeek, a new research question emerges: can LLMs perform causal reasoning at a level that could meaningfully replace human decision-makers, or should they remain human-led decision-support tools in high-stakes environments? This paper explores the causal reasoning capabilities of LLMs for operational and strategic military decisions. Unlike conventional AI models that rely primarily on correlation-based predictions, LLMs are now able to engage in multi-perspective reasoning, intervention analysis, and scenario-based assessments. We introduce a structured empirical evaluation framework to assess LLM performance through 10 de-identified real-world-inspired battle scenarios, ensuring models reason over provided inputs rather than memorized data. Critically, LLM outputs are systematically compared against a human expert baseline, composed of military officers across multiple ranks and years of operational experience. The evaluation focuses on precision, recall, causal reasoning depth, adaptability, and decision soundness. Our findings provide a rigorous comparative assessment of whether carefully prompted LLMs can assist, complement, or approach expert-level performance in military planning. While fully autonomous AI-led command remains premature, the results suggest that LLMs can offer valuable support in complex decision processes when integrated as part of hybrid human-AI decision-support frameworks. Since our evaluation directly tests this capability, this paradigm shift raises fundamental question: Is there a possibility to fully replace high-ranking officers/commanders in leading critical military operations, or should AI-driven tools remain as decision-support systems enhancing human-driven battlefield strategies?}, keywords = {}, pubstate = {published}, tppubtype = {article} } Military decision-making is inherently complex and highly critical, requiring commanders to assess multiple variables in real-time, anticipate second-order effects, and adapt strategies based on continuously evolving battlefield conditions. Traditional approaches rely on domain expertise, experience, and intuition, often supported by decision-support systems designed by military experts. With the rapid advancement of Large Language Models (LLMs) such as ChatGPT, Claude, and DeepSeek, a new research question emerges: can LLMs perform causal reasoning at a level that could meaningfully replace human decision-makers, or should they remain human-led decision-support tools in high-stakes environments? This paper explores the causal reasoning capabilities of LLMs for operational and strategic military decisions. Unlike conventional AI models that rely primarily on correlation-based predictions, LLMs are now able to engage in multi-perspective reasoning, intervention analysis, and scenario-based assessments. We introduce a structured empirical evaluation framework to assess LLM performance through 10 de-identified real-world-inspired battle scenarios, ensuring models reason over provided inputs rather than memorized data. Critically, LLM outputs are systematically compared against a human expert baseline, composed of military officers across multiple ranks and years of operational experience. The evaluation focuses on precision, recall, causal reasoning depth, adaptability, and decision soundness. Our findings provide a rigorous comparative assessment of whether carefully prompted LLMs can assist, complement, or approach expert-level performance in military planning. While fully autonomous AI-led command remains premature, the results suggest that LLMs can offer valuable support in complex decision processes when integrated as part of hybrid human-AI decision-support frameworks. Since our evaluation directly tests this capability, this paradigm shift raises fundamental question: Is there a possibility to fully replace high-ranking officers/commanders in leading critical military operations, or should AI-driven tools remain as decision-support systems enhancing human-driven battlefield strategies? |
| 19. | Theodore Tranos Nikolaos Fesakis, Thomas Vasileiou Sotirios Christopoulos Georgio Loukos Maria Koutsoupidou : AI-Based Energy Forecasting at Different Distribution Grid Levels to Support Baseline Definition and DSO Participation in LFMs. 2025 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe), IEEE, 2025, ISBN: 979-8-3315-2503-3. (Type: Conference | Abstract | Links | BibTeX) @conference{Tranos2025, title = {AI-Based Energy Forecasting at Different Distribution Grid Levels to Support Baseline Definition and DSO Participation in LFMs}, author = {Theodore Tranos, Nikolaos Fesakis, Thomas Vasileiou, Sotirios Christopoulos, Georgio Loukos, Maria Koutsoupidou}, url = {https://ieeexplore.ieee.org/abstract/document/11305676}, doi = {https://doi.org/10.1109/ISGTEurope64741.2025.11305676}, isbn = {979-8-3315-2503-3}, year = {2025}, date = {2025-10-20}, booktitle = {2025 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe)}, pages = {1-5}, publisher = {IEEE}, abstract = {A crucial aspect of Local Flexibility Markets (LFMs) is the definition of a baseline for energy production and demand forecasting, which serves as a reference for validating and compensating flexibility services. In this study, we explore the application of machine learning techniques, specifically Long Short-Term Memory (LSTM) networks, to establish accurate baselines for consumers and producers connected to the LV grid. The LSTM models leverage real historical demand and generation data from DSO smart meters in Mesogeia, Greece, combined with weather variables such as temperature and cloud coverage, to enhance forecasting accuracy. Our goal is to evaluate forecasting accuracy at the individual participant level and compare it with the accuracy obtained from forecasting on aggregated consumption/production data within a specific grid segment or using data from the secondary substation to which the participants are connected.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } A crucial aspect of Local Flexibility Markets (LFMs) is the definition of a baseline for energy production and demand forecasting, which serves as a reference for validating and compensating flexibility services. In this study, we explore the application of machine learning techniques, specifically Long Short-Term Memory (LSTM) networks, to establish accurate baselines for consumers and producers connected to the LV grid. The LSTM models leverage real historical demand and generation data from DSO smart meters in Mesogeia, Greece, combined with weather variables such as temperature and cloud coverage, to enhance forecasting accuracy. Our goal is to evaluate forecasting accuracy at the individual participant level and compare it with the accuracy obtained from forecasting on aggregated consumption/production data within a specific grid segment or using data from the secondary substation to which the participants are connected. |
| 20. | Asimina Dimara Konstantinos Kotis, Alexios Papaioannou Stamatis Chatzistamatis Nikolaos Evangeliou Chrysaphis Nathanailidis George Tsekouras : Data Collection, Organization, and Privacy-Preserving Preparation for Edge-Based LLMs in Legal Text Analytics. 5th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), 2025, ISBN: 979-8-3315-3556-8. (Type: Conference | Abstract | Links | BibTeX) @conference{Dimara2025, title = {Data Collection, Organization, and Privacy-Preserving Preparation for Edge-Based LLMs in Legal Text Analytics}, author = {Asimina Dimara, Konstantinos Kotis, Alexios Papaioannou, Stamatis Chatzistamatis, Nikolaos Evangeliou, Chrysaphis Nathanailidis, George Tsekouras}, url = {https://ieeexplore.ieee.org/abstract/document/11277858}, doi = {https://doi.org/10.1109/ICECCME64568.2025.11277858}, isbn = {979-8-3315-3556-8}, year = {2025}, date = {2025-10-16}, booktitle = {5th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME)}, abstract = {Providing fairness and privacy in automated legal text processing is an essential issue, especially with the increasing usage of Large Language Models (LLMs), in sensitive public sector applications. This paper presents a modular edge native domain-specific architecture for legal document processing that avoids cloud infrastructure and external APIs. The system combines local ingestion, semantic embedding, and retrievalaugmented generation to empower autonomous agents for applications such as bias detection and clause summarization. Inference is done exclusively on-device by a 4-bit quantized LLaMA model run by CPU-only runtimes. Tested on the CLEAR-Bias benchmark, the system gets 92% prompt relevance and 90% output coherence, inference latency below 6.5 s, and memory usage below 5.5 GB. These findings validate the effectiveness of privacy-preserving, regulation-conforming legal NLP in constrained environments.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Providing fairness and privacy in automated legal text processing is an essential issue, especially with the increasing usage of Large Language Models (LLMs), in sensitive public sector applications. This paper presents a modular edge native domain-specific architecture for legal document processing that avoids cloud infrastructure and external APIs. The system combines local ingestion, semantic embedding, and retrievalaugmented generation to empower autonomous agents for applications such as bias detection and clause summarization. Inference is done exclusively on-device by a 4-bit quantized LLaMA model run by CPU-only runtimes. Tested on the CLEAR-Bias benchmark, the system gets 92% prompt relevance and 90% output coherence, inference latency below 6.5 s, and memory usage below 5.5 GB. These findings validate the effectiveness of privacy-preserving, regulation-conforming legal NLP in constrained environments. |