Music Teaching Resource Recommendation with Multimodal Transformers

Multimodal semantic fragmentation and inaccurate user-intent modeling reduce the effectiveness of music teaching resource recommendation. To address these problems, this paper proposes MMT-MERec, a Multimodal Transformer for Music Educational Recommendation framework that integrates a multimodal Transformer with an educational knowledge graph. The method extracts 768-dimensional semantic vectors from audio, sheet music, video, and text through four pre-trained models: AST, MusicBERT, VideoMAE, and Sentence-BERT. A six-layer Cross-Modal Transformer encoder then performs fine-grained semantic fusion, using audio as a query to align the other modalities. A RotatE-embedded Skill Knowledge Graph containing 142 nodes constrains the recommendation loss to maintain teaching logic. A Skill-Aware Contrastive Learning objective based on user dynamics, including error rate and session duration, is used to optimize resource ranking. On an 8,200-sample dataset, MMT-MERec significantly outperforms the M3oE baseline, achieving HR@5 of 0.603 and NDCG@10 of 0.537. In cold-start scenarios, Cold-Start HR@5 reaches 0.401. These results show that the model jointly optimizes multimodal semantic alignment and educational knowledge constraints, improving recommendation relevance and teaching effectiveness.