Research And Implementation Of Music Recommendation Algorithm Based On Graph Of User Network

Recommendation algorithm is a very important part of online platform.In recent years,social recommendation algorithms mainly refer to various deep neural networks to extract deeper embedded representation of users and items,so as to better recommend items to users.However,in practical applications,especially in online music,there are some problems,such as sparse explicit social relations,static social relations,lack of comprehensive long-term and short-term interests of users,and imbalance of positive and negative samples.Aiming at the special problems of social music recommendation algorithm,this thesis studies the algorithms of implicit dynamic social graph generation,graph topology extraction and negative sampling,and realizes the music recommendation website based on the above algorithms.The main work of this thesis is as follows(1)In this thesis,a dynamic implicit social recommendation algorithm based on graph neural network is designed to solve the problems of sparse explicit social relationship,static user relationship and lack of comprehensive long-term and short-term interests in the practical application of social recommendation in music system.The algorithm extracts the dynamic implicit social graph from the sequential log through the nearest neighbor mechanism,extracts the topological structure information of the graph combined with the graph attention network,and deeply represents the fusion of spatio-temporal characteristics of users and items.The experimental results show that the accuracy of recommendation is effectively improved,and the MAE index reaches 0.3928 and AUC index reaches 0.7267 on the KKBox dataset.(2)In order to solve the problem that the negative samples are difficult to define and the positive and negative samples are unbalanced in the training samples of music recommendation algorithm,a negative sampling method in music recommendation algorithm is designed.In this method,popular negative sampling and random negative sampling are carried out for the bipartite graph of users and items.According to the business logic of music recommendation,different negative sampling methods are designed by using the node attributes of user’s social graph.The problem of unbalanced proportion of positive and negative samples in the training set is alleviated.The AUC index on KKBox dataset reaches 0.7361,which effectively improves the accuracy of recommendation.(3)In view of the special needs of music social recommendation,combined with the algorithm of this topic,a music recommendation website based on graph of user network is developed.The website can extract users’ dynamic social graph,and quickly return the results to recommend music to users through offline model training and online model loading.