A Hybrid Collaborative Filtering and Knowledge Graph-Based Cross-Domain Recommendation Method for Multimodal Teaching Resources

Synthesis: Liu, Sun, and Song (2026) propose a hybrid recommender that couples a pedagogically structured Knowledge Graph with behavior-aware collaborative filtering to personalize Multimodal AI English teaching-resource recommendation. Resources, concepts, and learner attributes are organized into four instructional dimensions — teaching context, cognitive level, technological feature, and cultural adaptability — whose user-dependent similarities are fused with a feature-based collaborative-filtering prediction through an ability- and progress-aware coefficient. On the English subset of the MARS dataset the model outperforms eight baselines (including the knowledge-graph-aware RippleNet and KGAT) on NDCG, hit rate, coverage, and cross-domain accuracy, and proves robust to partial modality removal.

Key Findings

• Four instructional dimensions structure the knowledge graph. Each resource entity is decomposed into teaching context, cognitive level, technological feature, and cultural adaptability; user-dependent attention weights over these dimensions let the model selectively emphasize or suppress each according to a learner's stated instructional requirements and behavior.
• Ability- and progress-aware fusion balances semantic and behavioral signals. An adaptive fusion coefficient λu rises with a learner's learning-ability index (LAE), progress-tracking index (LPT), and interest-preference index (IPI); higher-ability learners lean on personal behavioral patterns, while lower-ability or less-advanced learners get more structured guidance from the knowledge graph.
• Behavioral indices go beyond raw clicks. The interaction matrix aggregates click rate, view duration, and completion rate into an interest-preference index, and augments the learner feature vector with assessment-derived ability and progress proxies, reducing reliance on pure click frequency.
• Superior offline ranking and coverage. On the MARS English subset (4,800 users, 14,200 resources, 132,000 interactions; matrix sparsity 0.9981), CF–KG reaches NDCG 0.625 and HR 0.751, roughly a 7–8% NDCG gain over the best neural baseline, with coverage 0.736 and cross-domain accuracy 0.741 — the highest of all methods including RippleNet and KGAT.
• Modality robustness. Across text, video, and interactive resources the model holds NDCG 0.596–0.624 and HR 0.738–0.756; removing any single modality degrades NDCG by only 3.4–5.1%, and CF–KG under modality loss still beats the strongest full baselines.
• Both components matter. Component ablation (CF-only NDCG 0.584; KG-only 0.604) shows the two signals are complementary — removing the knowledge graph costs more (6.56% NDCG), but semantic relevance alone cannot capture individual behavioral preference.

Study Design & Method

The authors construct a knowledge graph whose entities are teaching resources, language concepts, skills, learner groups, and pedagogical attributes, with relation types grouped into content/domain, skill/goal, and pedagogical-alignment categories. Text resources are processed with TF–IDF to surface candidate topics, skills, and key concepts; each resource entity is then decomposed into four instructional-dimension vectors, and dimension-specific cosine similarities between resources are computed. On the behavioral side, an interaction matrix is enriched with the interest-preference, ability, and progress indices, and learner–learner similarity drives a feature-based collaborative-filtering prediction. A recommendation cycle starts from explicit instructional requirements (domain, skill, modality, cognitive level, cultural adaptation), performs k-hop semantic expansion over the knowledge graph to build a candidate set, refines candidates with collaborative filtering, and ranks them by the λu-weighted fusion of semantic and behavioral scores. Feedback updates the indices, fusion coefficient, and graph representations continuously.

What this means for practice

• Designers. Attach the four instructional dimensions — teaching context, cognitive level, technological feature, cultural adaptability — to every resource as structured metadata and rank on them, so recommendations reflect instructional fit rather than aggregated popularity and navigation generalizes across subject and modality boundaries.
• Designers. Gate behavioral influence on learner state through the ability, progress, and interest indices: let high-ability learners lean on their own interaction patterns while steering lower-ability or less advanced learners toward knowledge-graph guidance.
• Designers. Build the interaction matrix from completion rate and assessment-derived ability and progress proxies rather than click counts alone, since the interest-preference index combines click rate, view duration, and completion rate precisely because frequency under-describes preference, and lean on the knowledge-graph signal to offset the cold-start and data-sparsity problems that pure collaborative filtering faces.
• Researchers. Evaluate the pipeline against learning outcomes instead of ranking metrics: the authors state that the reported NDCG, hit rate, and coverage gains measure recommendation quality, not learning effectiveness.
• Researchers. Establish generalization beyond the English subset of MARS (4,800 users, 14,200 resources, 132,000 interactions) and report sensitivity analyses for the fusion parameters, which are tuned once on a validation split and then held fixed.

Limitations

• Offline evaluation only. All results are on historical interaction logs with ranking metrics (NDCG, HR, coverage, CDA); the authors state explicitly that reported improvements reflect recommendation quality, not learning effectiveness, and call for teacher assessments, learner studies, and outcome-based experiments as future work.
• English-subset scope. Evaluation uses the English subset of MARS, so generalization to other languages, domains, or datasets is untested.
• Model-learned vs. empirically fixed weights. IPI aggregation weights and fusion parameters are tuned on a validation split and held fixed; no sensitivity analysis is reported for them.
• No online/cold-start deployment test. Robustness is tested by removing modalities, not by evaluating truly cold-start users or live personalization.

Citation

Liu, B., Sun, X., & Song, J. (2026). A Hybrid Collaborative Filtering and Knowledge Graph-Based Cross-Domain Recommendation Method for Multimodal Teaching Resources. Computers and Education: Artificial Intelligence.