VMF-GCN: A music recommendation model integrating Variational Matrix Factorization and Graph Convolutional Networks

With the rapid development of digital streaming platforms, music recommendation systems have become increasingly important. However, existing methods face significant challenges in handling the inherent uncertainty and noise in user behavior data and often fail to effectively capture high-order interactions between users and music items. To address these limitations, this paper proposes a music recommendation model, VMF-GCN, which integrates Variational Matrix Factorization and Graph Convolutional Networks within a dual-branch fusion architecture. Specifically, the variational matrix factorization branch employs a variational autoencoder to probabilistically model user interaction behaviors, thereby capturing preference uncertainty and mitigating the impact of noise on embedding representations. Meanwhile, the graph convolutional network branch constructs a user–music interaction graph and performs message propagation to learn deep embeddings that encode high-order neighborhood information. A score fusion mechanism is then used to integrate the predictions from both branches. Extensive experiments conducted under a unified 5-fold cross-validation protocol demonstrate that VMF-GCN consistently outperforms six representative baseline methods, including Neural Collaborative Filtering (NCF) and the state-of-the-art graph-based model LightGCN. In particular, VMF-GCN achieves an average MAE of 0.4775 ± 0.0250 and an average MSE of 0.4559 ± 0.0334, yielding the best performance among all compared methods. Further ablation studies and parameter sensitivity analyses confirm the effectiveness of each model component and the robustness of the proposed approach. These results indicate that VMF-GCN provides an effective and robust solution for music recommendation tasks.