Dynamic Node Graph Neural Network for Multimodal Music Recommendation

Abstract

Music recommendation systems are becoming increasingly popular among users. With the explosive growth of songs on the web, most music streaming platforms have launched online music listening services, providing millions of music choices for users. However, how to accurately recommend songs for users that match their preferences has become a challenging challenge, which we call music cold-start matching. In this paper, we delve into the multimodal information of music and take full advantage of the unique strengths of graph neural networks in capturing the collaborative filtering relationship between users and music. However, due to the inherent characteristics of graph neural networks, it is difficult to easily add new nodes to perform subsequent tasks in the inference phase. Therefore, we creatively propose a novel neural network architecture, the dynamic node graph neural network. In the training phase, we adopt a knowledge distillation strategy, using the graph neural network as the teacher model and the dynamic node graph neural network as the student model, thus enabling the student model to comprehensively learn and master the collaborative filtering relationship between users and music. In the inference phase, we use the trained dynamic node graph neural network to match new music accurately. After extensive experimental validation on the MSD public dataset, our approach demonstrates excellent effectiveness and efficiency, bringing users a more accurate music recommendation experience. Experimental results show that the state-of-the-art method improves our model by an average of 11.8% on the complete dataset and 7.1% on the cold-start problem compared to the best method, proving the effectiveness of our model.