Research On Music Recommendation Algorithm Based On Heterogeneous Information Network And User Preference

Recommendation system has a subtle impact on people’s life,for different users to bring better network service experience.However,with the explosive growth of information,traditional recommendation algorithms face many tests,such as difficult to model social relations,heterogeneous information networks and other structures,and are also limited by users’ cold start and sparse data.In addition,traditional recommendation algorithms are also faced with the lack of recommendation diversity,poor interpretability and other problems.Heterogeneous information network has great advantages in processing graph structure data and can capture rich semantic information between entities.In recent years,the application of heterogeneous information network to recommendation algorithm has attracted much attention.In this paper,the existing heterogeneous information network recommendation is improved by incorporating attention mechanism,and the long-term and short-term preferences are modeled respectively in the heterogeneous information network considering the short-term interests of users.The main work is as follows:A heterogeneous information online music recommendation algorithm(MCHAN)integrating attention mechanism is proposed.Firstly,MAGNN algorithm is used to embed heterogeneous graph to obtain the global user feature embedding vector.As the local preferences of users are often rich and diverse,MCHAN uses the local conditional subgraph attention network layer to distill the heterogeneous subgraph and selects the second-order neighbors of nodes,taking the importance of nodes into account.Finally,the conditional attention mechanism is used to extract the local preferences of users and make top-N recommendation.At the same time,considering the influence of short-term user preferences and proposes a fusion of both short-term and long-term user preferences(MCHAN＿ST),music recommendation algorithm based on MCHAN algorithm is used to capture the user’s preference for a long time,the door control cycle unit(GRU helped)network short-term preference of capture user personalized recommendation of short-and long-term user preferences to the user.Two evaluation indexes were selected and NDCG for testing,and analysis of experimental results on the Last.fm data set showed that:MCHAN algorithm improved 9.55%and 8.88% respectively in the above two evaluation indicators by comparing with the baseline model.MCHAN＿ST algorithm also improved the results in NDCG and MCHAN＿ST algorithm to varying degrees by incorporating users’ short-term preferences,indicating that the extraction of users’ short-term preferences is of certain help to improve the accuracy of the recommendation results.