Personalized music recommendation algorithm based on machine learningPersonalized music recommendation algorithm based on machine learningL. Liu et al.

Music is a joint leisure and entertainment activity in people’s daily lives. However, with the intensification of the problem of information overload, it has become more difficult for users to quickly find the songs they are interested in from the massive amount of music. Currently, many initiatives are being used to help users rapidly match the target songs, and none other than the music recommendation technology. On the one hand, this paper designs a recall model LSDM (Long Short-term Deep Matching Model) based on users’ long-term and short-term preferences to help users quickly and accurately identify a small range of candidate songs that may be of interest from massive data. Firstly, the user behaviour sequence is divided into long-term and short-term sequences, and the long-term and short-term interest preference representations of users are obtained by modelling the long-term and short-term sequences, respectively. Then, the user representation vector was generated by fusing the user’s long-term and short-term preferences through the gating mechanism, and a candidate recommendation list was generated for each user based on the similarity between the user vector and the song vector. On the other hand, this paper designs an Improved Deep Session Preference Model (IDSPM) to sort the candidate list further to recommend songs to users more accurately. The ranking model divides the behaviour sequence into several sessions and uses Transformer and BiMO-LSTM to model the sequence within and between sessions, respectively, to accurately obtain the expression of interest in each session and effectively capture the evolution law of session interest. Then, based on the cosine similarity and Euclidean distance between vectors, a new attention weight calculation method is designed, and the weighted sum of the conversational interest vectors is performed to obtain the user’s interest characteristics. The user’s interest characteristics are input into the DNN, and other characteristics are used to predict the user’s click probability of the candidate song. The final recommendation list is generated according to the click probability. Finally, under the #\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\#$$\end{document}nowplaying-RS dataset, simulation experiments, comparative experiments and ablation experiments are carried out on the two proposed algorithm models, which fully confirm the effectiveness of the models. In addition, a hyperparameter sensitivity experiment is carried out to analyze the influence of crucial hyperparameters on the algorithm model.