MFF-CF: a multi-feature fused collaborative filtering framework for personalized learning resource recommendation in smart education

Abstract

Personalized learning resource recommendation in smart education platforms faces three persistent challenges: severe interaction matrix sparsity, cold-start failure for newly introduced resources, and static modeling of temporally evolving learner preferences. This paper proposes MFF-CF (Multi-Feature Fused Collaborative Filtering), a unified recommendation framework integrating three domain-specific modules. The Hybrid Similarity Computation (HSC) module fuses bidirectional GRU-encoded behavioral sequence embeddings with adjusted cosine similarity to enrich the user similarity space under sparse conditions. The Knowledge Graph-Augmented Item Representation (KGAIR) module applies a relational graph convolutional network over an educational knowledge graph encoding prerequisite relationships and concept associations, enabling non-zero recommendation precision for cold-start resources absent from training history. The Dynamic Temporal Interest Decay (DTID) module models learner preference evolution via learner-specific exponential decay rates, ensuring recommendations reflect current rather than time-averaged interests. Evaluated on MOOC-Edu and EdNet against six baselines including UserCF, ItemCF, MF, NCF, LightGCN, and BERT4Rec, MFF-CF achieves statistically significant improvements of 8.3% in Precision@10, 11.7% in Recall@10, and 9.4% in NDCG@10 over the strongest baseline on EdNet (p < 0.05), and is the only method to produce non-zero cold-start Precision@10 (0.0387 on MOOC-Edu; 0.0312 on EdNet). Ablation experiments confirm independent and complementary contributions from each module. These results demonstrate that MFF-CF provides a practical and scalable solution for intelligent educational resource delivery.