EKM 2026 Abstracts

Area 1 - Educational Knowledge Management
Full Papers Short Papers

Area 1 - Educational Knowledge Management

Full Papers
Paper Nr: 5
Title:

Towards Shareable Gamification Knowledge: An Ontological Approach to Support Teachers in Gamified E-Learning Systems
Authors:

Yara Gomaa, Christine Lahoud, Marie-Hélène Abel and Sherin Moussa
Abstract: Gamified e-learning systems increase enhanced learner engagement and reduce dropout rates. However, gamification strategies are often embedded at the system level and applied uniformly across different learning contexts, with limited consideration of the teachers’ intentions and expertise. As a result, the teacher’s role in selecting and sharing gamification strategies remain underrepresented in existing models and gamified-learning approaches. To address this gap, this paper proposes the teacher in gamified e-learning context (TGC) ontology, a modular ontology that provides a shared semantic representation of teacher-related knowledge, collaboration concepts, and gamification elements in e-learning environments. TGC is designed to support teacher collaboration by enabling the explicit representation and exchange of gamification experiences, resources, and expertise according to teachers’ objectives, gamified teaching situations, and expertise level. The ontology was developed following the NeOn methodology and evaluated through (1)reasoning-based verification, (2) SPARQL querying based on competency questions,(3) quantitively through OntoMetrics schema analysis, and (4) automatic evaluation for TGC alignment to FAIR principles. The results indicate that TGC is logically consistent, accessible, structurally expressive, and reusable as a semantic foundation for teacher-centered gamified e-learning applications.
Download
Paper Nr: 6
Title:

Ontology-Based Generative AI Personalization in Game-Based Learning
Authors:

Ameny Rjiba, Lilia Cheniti Belcadhi and Judita Kasperiuniene
Abstract: Generative Artificial Intelligence is increasingly used to enable personalization in digital learning environments, yet its role within game-based learning remains insufficiently mapped and conceptually fragmented. Existing research primarily focuses on personalization at the individual level, neglecting group-level support, adaptive collaboration, and role-based guidance in shared learning environments. To clarify the structure of this emerging field, this study uses a science-mapping analysis to examine the intersection of generative AI, personalization, and game-based learning. Using a structured Scopus search and established science-mapping techniques, the analysis identified dominant conceptual themes and thematic developments. The results reveal significant gaps related to adaptive support, explainable feedback, and collaborative personalization mechanisms. Guided by these insights, the paper proposes an ontology-based framework that extends personalization to collaborative, game-based settings through role prompts, micro-scenarios, and adaptive explanation mechanisms. The framework is accompanied by an implementation that demonstrates its key components and illustrates its applicability within a game-based learning environment. This study provides a structured map of the field and lays the groundwork for future system design.
Download
Paper Nr: 10
Title:

xCARE: Explainable Continual Adaptation and Remediation Engine via Agentic RAG for Concept Drift in Intelligent Tutoring Systems
Authors:

Wissal Loussaief and Nacim Yanes
Abstract: Intelligent Tutoring Systems (ITS) often lose predictive accuracy when students change their learning behavior or engagement, a problem known as concept drift. Traditional adaptation methods updating these models frequently destroy previously learned knowledge—an issue called catastrophic forgetting—and act as opaque ”black boxes” providing no pedagogical explanation. In this paper, we introduce xCARE (eXplainable Continual Adaptation and Remediation Engine), a Human-in-the-Loop Decision Support System solving this problem through two main contributions. First, xCARE uses a Continual-LoRA router to rapidly update an Extended Long Short-Term Memory (xLSTM) predictive network, adapting to new behaviors while preserving past knowledge. Second, it features an Agentic Retrieval-Augmented Generation (RAG) module translating raw statistical drift alerts into explainable, gamified remediation tasks for teachers to validate. Evaluated on the ASSISTments dataset, xCARE demonstrated high adaptive efficiency, achieving a +24.50% gain in Accuracy and +26.75% in AUC-ROC during a sudden concept drift. Furthermore, an ablation study confirmed our architecture maintained a perfect Backward Transfer (BWT) of +0.00%, eliminating the severe -20.47% memory loss observed in standard full fine-tuning. This ensures both algorithmic stability and practical pedagogical usability.
Download
Paper Nr: 11
Title:

x_GATES: xLSTM Gamified Alerts for Tracing and Explaining Shifts in Intelligent Tutoring Systems
Authors:

Oumaima Méchi and Nacim Yanes
Abstract: Most Intelligent Tutoring Systems (ITS) operate under a static environment assumption, failing to account for the non-stationary evolution of students’ cognitive strategies. This research addresses this gap by introducing xGATES, a human-centric cybernetic framework designed for proactive pedagogical monitoring and metacognitive support. Operating as a continuous feedback loop, xGATES synchronizes a stabilized xLSTM predictive core with a bimodal meta-detection engine. This architecture concurrently monitors algorithmic residuals and latent engagement telemetry derived from gamified scoring systems to identify concept drift before persistent academic failure occurs. Evaluated on the ASSISTments benchmark, the predictive core achieves an AUCROC of 0.9009. Crucially, the bimodal engine identifies cognitive disengagement prior to a 27.1% collapse in model reliability, while a hierarchical statistical audit (Cohen’s h > 0.8) successfully filters transient noise to ensure intervention precision. By translating abstract statistical anomalies into natural-language diagnostics via a Generative XAI layer, the system initiates a “Tactical Pause”—a data-driven reflection period that empowers learners to self-regulate. xGATES thus transforms algorithmic failures into transparent pedagogical opportunities, fostering resilient and autonomous learning environments.
Download
Short Papers
Paper Nr: 8
Title:

Resource Recommendations for Teachers: An Approach Based on Technical Skills
Authors:

Nader N. Nashed, Elsa Negre and Marie-Hélène Abel
Abstract: The rapid evolution of educational technologies has significantly transformed teaching practices. Teachers are increasingly required to integrate a wide range of digital tools into their courses, from simple programming environments to advanced artificial intelligence systems. While these technologies create new pedagogical opportunities, they also increase the technical demands placed on teachers. Existing educational resource recommender systems mainly focus on disciplinary or pedagogical aspects, often neglecting the technical skills required to effectively use these resources. As a result, teachers may receive recommendations that are pedagogically relevant but technically difficult to implement. In this paper, we propose a constraint-based recommender system designed to support teachers in selecting pedagogical resources according to both their scientific and technical competencies. The approach relies on an ontology-based representation of teachers and pedagogical resources, enabling the formalization of compatibility constraints between resource requirements and teacher skills. The recommendation process is formulated as a constraint satisfaction problem in which resources are selected based on their compatibility with the teacher profile and the teaching environment. A prototype architecture is presented together with a recommendation model integrating semantic constraints derived from the ontology. This approach aims to provide more realistic and usable recommendations for teachers by explicitly considering the technical dimension of educational resources.
Download
Paper Nr: 9
Title:

To Fail or Not to Fail? A Multi-Dataset Benchmarking Analysis for Dropout Prediction
Authors:

Etienne Wagner, Cheikh Brahim El Vaigh, Manel Addi, Christophe Nicolle and Kokou Yetongnon
Abstract: Identification of trainees at risk of failure is crucial for implementing timely and effective support within professional training platforms. This study presents a comprehensive benchmarking analysis of trainee failure prediction across three distinct datasets, incorporating demographic information, behavior features and course outcomes. Our methodology encompasses extensive feature engineering, class imbalance handling through SMOTE, and systematic hyperparameter optimization via Optuna. This pipeline was applied to a broad suite of classical machine learning and ensemble boosting approaches. Experimental results demonstrate that well-tuned methods consistently achieve strong predictive performance, with the best models reaching an F1-score of up to 99% on our professional e-learning dataset. The results highlight the effectiveness of combining robust preprocessing pipelines with automated optimization strategies as a reliable and interpretable framework for trainee dropout prediction. These findings provide practical guidance for instructors and learning system managers looking for scalable data-driven solutions to support at-risk learners in professional training settings. The use of SHAP with our best model opens the path for an explainable adaptive learning system that can be easily managed by educational experts.
Download
Paper Nr: 12
Title:

KG-BAL: Knowledge Graph-Guided Bayesian Active Learning for Conversational Document Retrieval
Authors:

Anas Ouhannou, Yazan Mualla, Christine Lahoud and Sana Nouzri
Abstract: When users issue ambiguous queries against large document collections, standard retrieval systems either return overwhelming result lists or make a single best-guess. Both strategies shift the disambiguation burden to the user. This challenge arises in many information-seeking contexts: citizens navigating government procedures, students searching for courses matching their objectives and level, or professionals discovering relevant technical documentation. We propose KG-BAL, a framework that combines three complementary techniques to address this problem in a principled way: (i) a domain Knowledge Graph (KG) encoded as an OWL ontology that provides the structured attribute space over which clarification dialogue operates; (ii) a Bayesian belief tracker that maintains a probability distribution over candidate documents and updates it after each user response; and (iii) Expected Information Gain (EIG), an information-theoretic criterion that selects the question producing the largest expected reduction in uncertainty. We evaluate KG-BAL on a corpus of 614 Moroccan Arabic administrative procedures and show it outperforms traditional retrieval and LLM-based approaches.
Download