MuseFuse: Multi-View Feature Fusion for Music Emotion Recognition with Prototype-Guided Alignment and Mutual Distillation

Music emotion recognition supports affective computing in interactive art, games and creative tools by mapping musical content to perceived emotional states. Multi view audio representations such as Mel spectrograms and cochleagrams provide complementary information but suffer from misaligned feature spaces, noisy pseudo labels and inconsistent predictions across branches. This work presents the MuseFuse framework for music emotion recognition, which employs dual view encoders, adaptive gated fusion and three classification heads to obtain stable affective predictions. Prototype Guided Alignment uses a shared prototype bank to align Mel, cochleagram and fusion embeddings in a semantic space, while Reliability Guided Self Training assigns reliability weighted pseudo labels to reduce noise in unlabeled data. A Tri Branch Mutual Distillation module encourages consistent predictions among Mel, cochleagram and fusion heads through symmetric knowledge transfer. The joint optimisation of classification, domain alignment, consistency, prototype, pseudo label and distillation objectives produces robust fusion representations for emotion recognition. Experiments on multiple benchmark datasets including Memo2496, PMEmo, DEAM and 1000songs show that MuseFuse consistently outperforms strong single view and multi view baselines in accuracy, F1 and AUC, and analyses confirm improved cross view alignment, reliability under noisy labels and generalisation across datasets.