Fair and explainable educational recommendations with a hybrid Graph-GRU framework

Synthesis: Fair and explainable educational recommendations with a hybrid Graph-GRU framework

Key Findings

• The paper introduces the Hybrid HKG-GRU framework, combining heterogeneous Knowledge Graph embeddings with sequential (GRU) modeling to capture both the relational structure of course materials and the temporal dynamics of learner interactions.
• Evaluation on Moodle LMS logs from 152 students, 59 resources, and approximately 150k interactions achieved strong predictive performance with HR@10 = 0.68 and MRR = 0.41.
• The framework integrates three responsibility-oriented contributions: multi-objective training with Group Distributionally Robust Optimization (GroupDRO) for fairness, Maximum Marginal Relevance (MMR) reranking to reshape exposure patterns, and built-in model-centric explainability through path-based and counterfactual analyses.
• Results showed high intra-list diversity and moderate catalog coverage, with moderate counterfactual stability for many learners (median CR@10 = 1.0), although catalog-level popularity bias remained evident.
• The work addresses popularity bias and cold-start fairness in educational recommenders, where students with limited participation histories risk receiving less reliable support while popular resources dominate recommendation lists.

Study Design & Method

The study targets a known failure of accuracy-focused recommenders: students with limited participation histories receive less reliable support, while highly popular resources may dominate lists and limit access to other useful materials. The hybrid architecture addresses this by modeling course-material structure and learner sequences jointly, then applying GroupDRO training to make performance robust across learner groups, MMR reranking to diversify exposure, and path-based and counterfactual explainability to make recommendations auditable.

What this means for practice

• Software developers. Build fairness, diversity, and explainability into the model as training-time objectives rather than post-hoc reporting: combining GroupDRO training, MMR reranking, and path-based and counterfactual explanations reached HR@10 = 0.68 and MRR = 0.41 on the same interaction logs, showing that responsibility objectives need not cost predictive accuracy.
• Instrument exposure inequality as a standing metric, not a one-off audit. Catalog-level popularity bias remained evident after GroupDRO and exposure-based regularization, so track catalog coverage and Gini exposure on every retrain rather than declaring Bias Mitigation complete.
• Design explicitly for cold-start and low-activity learners, who risk receiving less reliable support because fairness here was defined as robustness across participation-based cohorts of learner activity.
• Keep a human in the loop on the navigation decision. The path-based and counterfactual explanations are model-centric, so present them to instructors and learners as reviewable evidence consistent with Human-in-the-Loop oversight, not as final determinations.
• Do not read activity level as ability or need when acting on recommendations — the public Moodle logs contain no achievement, Prior Knowledge, learner profile, or demographic attributes, so a low-activity learner may be struggling, disengaged, or already familiar with the material.

Limitations

• Dataset scale and diversity: evaluation rests on a single course with limited institutional and demographic diversity, constraining generalizability of the fairness and robustness findings.
• Scope of the fairness definition: fairness is operationalized through participation-based cohorts (learner activity level) because the public Moodle dataset lacks achievement, prior knowledge, learning profiles, or demographic attributes; the results are therefore an audit of behavior across engagement levels rather than a full assessment of educational equity. As the authors note, a low-activity learner may be struggling, disengaged, or already familiar with the material, and activity level does not necessarily reflect ability or need.
• No human-centered bias auditing: no instructor/student evaluation was conducted, so interpretability and trust claims remain model-centric and technical.
• What the evaluation does not capture: only short-range item→precedes→item edges within a single course are modeled, so multi-semester trajectories, cross-course relationships and contextual shifts are out of scope; and substantial catalog-level popularity bias persists despite GroupDRO and exposure-based regularization, as the high Gini exposure metrics show.

Citation

Evangelista, E., & Bukhari, S. M. S. (2026). Fair and explainable educational recommendations with a hybrid Graph-GRU framework.