Intelligent tutoring in dynamic domains: a graph-based system for comparative analysis of adaptive algorithms

Synthesis: Csépányi-Fürjes and Kovács implement and empirically validate G4L, a graph-based intelligent tutoring system for fast-changing domains. Built on the Evolving Knowledge Space Graph (EKSG) model (Neo4j) plus a relational database (MySQL) for real-time tracking, it uses intuitionistic fuzzy logic (IFL) to represent learner knowledge states and forgetting, and a generative-AI assistant to auto-generate knowledge units, prerequisite relations, and quiz questions (teacher-in-the-loop review). Three adaptive algorithms — Bayesian Network (BN), Knowledge Space Theory (KST), and the proposed Weighted Distance Dependent Induction (WDDI) — are compared in a 45-participant classroom study. The Bayesian algorithm produced the highest knowledge gains (0.717 → 0.887, +24%), while log analysis revealed self-regulated learning dominated system recommendations.

Core Finding

A hybrid ITS that combines a structured graph-based domain model (EKSG), a relational database for real-time activity/knowledge tracking, and a generative-AI content-creation assistant is feasible and pedagogically effective in a dynamic domain. In the adaptive algorithm comparison, the Bayesian knowledge propagation algorithm yielded the highest measured learning gains (BN +24% vs KST +5%, WDDI +1%), and it also had the lowest prediction error alongside a graph-neural-network baseline — supporting it as the strongest adaptive algorithm in this coarse-grained curriculum setting.

What the System Is

G4L operationalizes the EKSG model — a graph-based knowledge representation extending Knowledge Space Theory and Competence-Based KST — in a working ITS:

• EKSG graph (Neo4j): directed graph of Knowledge Unit, Material Unit, and Test Unit nodes linked by prerequisite_of edges, making evolving domain structure explicitly visible to learners.
• Relational database (MySQL): learner activity logs, knowledge state data, system config, and teacher data for real-time tracking and adaptive decisions.
• Learner subsystem + Teacher subsystem + Knowledge retention subsystem: scheduling forgetting updates so knowledge decays realistically over time.
• Generative-AI assistant: auto-generates knowledge units, prerequisite relations, and quiz items from existing documentation, with human expert refinement (teacher-in-the-loop).

How Adaptation Works

• Learner assessment responses are mapped to IFL mastery values per knowledge unit (truth/falsity/hesitation on an adapted Atanassov IFL triangle).
• These observed values are propagated through the knowledge graph using one of three algorithms (BN, KST, or WDDI) to estimate mastery of related (unassessed) units.
• The system recommends the next knowledge unit: units outside the "KNOWS" region are ranked by proximity to the "knows" vertex.
• Forgetting is modeled as a vector shift toward "unknown" using the Ebbinghaus forgetting curve, driven by elapsed time and repetitions; units most vulnerable to decay are prioritized for review.
• Learners may follow or override recommendations, supporting Learner Agency and self-regulated learning.

Results

• RQ1 (structure/navigation): Knowledge Decomposition and Graph-based Navigation Usability rated highest (avg 4.41, rubric level 3); Prerequisite Logic Clarity lowest (4.16).
• RQ2 (algorithms): measured knowledge improved in all groups — BN 0.717→0.887 (+24%, score 3), KST 0.919→0.967 (+5%, score 2), WDDI 0.871→0.882 (+1%, score 2). Final rubric scores BN 3, KST 2, WDDI 2. In a Simulation, BN (error 14) and a GNN baseline (20) tracked true IFL truth values best vs KST (94), WDDI (62), uniform static (100).
• RQ3 (tracking): system scored "Good (3)" (11/13 tracking features, 85%). BN group logged the most per-learner records (855.9) and the most active test engagement; the most active learner (112 interactions) was in the KST group.
• Self-AI Regulation in Education: most learners disregarded system recommendations (learn-followed 18–32%, repeat-followed 7–9%), indicating their own decisions took precedence over the suggested path.

Key Strengths and Caveats

• Strengths: combines structured domain modeling, fine-grained real-time activity logging (13 event types), adaptive content delivery, and generative content creation; teacher-in-the-loop keeps Pedagogies and Teaching Strategies under human control; log-based tracking is lightweight and non-intrusive.
• Caveats: small sample (45, 3 dropouts, 28 questionnaire respondents); single subdomain (JDBC API, 15 KUs); short study window (limited long-term retention insight); no control group or alternative-ITS comparison; KST predicted-mastery values stuck at 1.000 (full-mastery inference) and should be read with caution; WDDI's proposed bi-directional propagation underperformed on measured gains.

Relevance to the knowledge base

This is a development article that demonstrates a complete, implemented pipeline from a knowledge-graph domain model through adaptive algorithms and real-time learning analytics to generative content creation — an integration rarely reported end-to-end. Its empirical comparison of multiple adaptive algorithms (with the Bayesian algorithm showing the highest knowledge gains) is directly relevant to comparative ITS design, while its explicit modeling of forgetting, IFL-based uncertainty, and log-driven self-regulated-learning analysis connects several core knowledge base themes. The GenAI content-generation component with teacher-in-the-loop validation offers an applied example of generative AI in instructional design.

What this means for practice

• Designers. Keep a teacher in the loop when GenAI generates course content: G4L's assistant drafted knowledge units, prerequisite relations, and quiz items from existing documentation, but every item passed through expert refinement before use.
• Designers. Prefer Bayesian knowledge propagation over KST and WDDI for coarse-grained curricula — it delivered +24% measured knowledge (0.717 → 0.887) against +5% (KST) and +1% (WDDI), and tracked true mastery values with the lowest error (14 vs 20 for a graph-neural-network baseline and 94 for KST).
• Designers. Model forgetting explicitly, as G4L does with an Ebbinghaus decay driven by elapsed time and repetitions, and schedule review of the units most vulnerable to decay instead of relying on the most recent score.
• Designers. Let learners override recommendations: in this study only 18–32% of learn-recommendations and 7–9% of repeat-recommendations were followed, so learners treat the suggested path as advice rather than a mandate.
• Researchers. Verify that an algorithm's internal mastery estimates match observed performance before reporting gains — KST's predicted-mastery values saturated at 1.000 here, which the authors flag as warranting caution.

Limitations

• The sample was 45 participants with 3 dropouts, and only 28 learners completed the post-intervention questionnaire, which the authors state may limit the representativeness of the results.
• Instructional content covered a single subdomain of just 15 knowledge units (the JDBC API), restricting generalizability to broader curricula.
• The study ran within a single semester-length window, which the authors state is too short to reveal long-term knowledge retention.
• The system was not tested against a control group or an alternative ITS platform, so the BN/KST/WDDI comparison is internal to G4L and cannot establish its advantage over other tutoring systems; knowledge outcomes rest on mastery ratios (correct answers over attempts) and log data, which the authors note cannot capture cognitive states, motivation, or off-platform learning and give no causal explanation for performance differences.

Citation

Csépányi-Fürjes, L., & Kovács, L. (2026). Intelligent tutoring in dynamic domains: a graph-based system for comparative analysis of adaptive algorithms with intuitionistic fuzzy logic and forgetting. Educational Technology Research and Development.