Online Music Recommendation Algorithm Based on Heterogeneous Neural Network

Traditional recommendation algorithms mainly rely on historical interaction information between users and music projects, but ignore the collaborative signals in user interactions, and are limited by the sparsity of interaction data. To solve these problems, this paper combines heterogeneous graph neural networks and designs the HNMRA algorithm by splitting the graph structure, decomposing the complex network into multiple focused on specific relationships subgraphs. By using the user's explicit feedback, such as user comments, we need to calculate preference weights. On each subgraph, the GNN learns the user and music attributes and behavior preference representations by using the attention mechanism to fuse information from different subgraphs, generating more precise user and music embedding representations. Based on these embedding representations, the top-K recommendation is performed, and the K most likely interesting music is recommended to the user. The advantage of the algorithm is that it can handle multidimensional relationships simultaneously, improve the recommendation accuracy by using explicit feedback, and achieve personalized preference modeling. To solve the challenges of computational complexity and data sparsity, this paper proposes optimizing the GNN structure design, adopting efficient graph partitioning and parallel processing technologies, and combining implicit feedback, such as playback times, to enhance model robustness. Experimental results show that the proposed research not only improves the comprehensiveness and accuracy of recommendations, but also provides personalized recommendation results, which has high research value and application potential.