Collaborative filtering algorithm optimization and user personalized modeling in music recommendation systems

Abstract

As music gradually becomes an indispensable part of people’s lives, the optimization of music recommendation systems has become a research hotspot. In response to the problems of poor recommendation accuracy and low degree of personalization in music recommendation systems, this article combined the optimized LGCN-A (Light Graph Convolution Network-Attitude) algorithm to conduct personalized modeling and recommendation research in music recommendation systems. Firstly, the collected data is cleaned and normalized, and feature data of users and music are extracted. Then, a graph structure between users and music is constructed, and each user and music node is initialized with a feature vector operation. The original feature vectors of users and music can be mapped to the embedding space through linear transformation, and the embedding vector can be initialized using Gaussian distribution and regularized. Finally, based on the traditional NGCF (Neural Graph Collaborative Filtering) algorithm, lightweight processing can be achieved by reducing nonlinear activation functions and feature transformation steps, and self-attention mechanism can be introduced to assign different weights to different users and music. The experiment is based on the public dataset Million Song Dataset to predict the interaction between users and music, and generate a recommendation list. The results show that at k of 20, the recommendation hit rate of the LGCN-A algorithm reached 0.95, which is 0.19 higher than the traditional NGCF algorithm, and the overlap is only 21.43%. The LGCN-A algorithm has improved the accuracy of recommendations in music recommendation systems, ensuring different levels of user personalization and providing strong support for the further development of recommendation systems.