A Dual Modeling Framework for Music Recommendation Via GraphSAGE and Deep Interest Networks (DIN)

Abstract

This study proposes a music recommendation system that combines the GraphSAGE and DIN algorithms to improve the accuracy and personalization of recommendations. Using a dataset of 50,000 users, 100,000 songs, and 3 million listening records, the system evaluates performance using metrics like AUC and HitRate. The GraphSAGE algorithm extracts users' long-term interest preferences from heterogeneous user-song graphs, while DIN dynamically adjusts users' short-term interests through attention mechanisms. Experimental results show that the hybrid model outperforms standalone GraphSAGE and DIN, with a recommendation accuracy of 92%, which is 7% higher than GraphSAGE and 4% higher than DIN. The model also shows significant improvements in user satisfaction (4.5 vs. 4.1 for GraphSAGE and 4.3 for DIN), demonstrating its effectiveness in handling large-scale music recommendation tasks with dynamic user preferences. Through this dual modeling method of structure + dynamics, this study effectively makes up for the shortcomings of a single algorithm in dealing with complex user behaviors. For the experiment of recommendation coverage, the GraphSAGE algorithm shows a higher recommendation coverage, especially when the number of music exceeds 100, the recommendation coverage is stable at around 75%, which is significantly better than DIN's 65%. At the same time, this paper also analyzes the relationship between system response time and performance. Experimental results show that GraphSAGE can maintain high recommendation accuracy when the response time is short, while DIN achieves the best performance when the response time is 10 hours. The performance advantages of the two in different situations complement each other. The research in this paper not only proposes a new music recommendation system architecture, but also proves through experiments that the effective combination of GraphSAGE and DIN can significantly improve the accuracy, coverage and user satisfaction of the recommendation system. This research provides a new idea and technical path for the application of recommendation system in complex data and multi-dimensional interest modeling scenarios.