Research And Implementation Of Music Recommendation System Based On User Behavior And Audio Features

With the advent of the era of big data and 5G,the amount of data on the Internet continues to grow,and the problem of data overload is becoming more and more serious.In recent years,with the continuous development of big data and artificial intelligence technology,recommendation systems have been used in news websites,short video platforms,online music and other industries as an effective method to solve the problem of data overload.For music recommendation field,the previous workers did a lot of research,but there are less real-time traditional methods can not achieve personalized recommendations and new users may exist "cold start" and so on.Aiming at the shortcomings of traditional music recommendation systems,this article makes some improvements and explorations to provide users with accurate music that they are interested in.This paper refers to the traditional music recommendation method and proposes a personalized recommendation method combining user behavior and music audio.The main research work of this thesis is:(1)First,process the data set.Divide the data set into two subsets with different degrees of sparseness.The implicit scores such as the number of listening times,likes,and favorites are converted into explicit scores of 0-15,and then relevant tools are used to extract music audio features.(2)Secondly,a group recommendation algorithm based on user community is proposed.Firstly,a user-music bipartite network is constructed,which combines user history listening data and user information to calculate interest preferences,and divides them into communities.Users with the same music preferences are divided into the same community.For the user,determine the user's community,through the community-based collaborative filtering model,the user can calculate the preference score,thereby recommending the Top-N music with the highest score.This article also implements a group recommendation algorithm based on the user community on the Spark computing platform,which improves the recommendation efficiency.(3)Then,referring to the related research on image classification,the idea of Relation-Learning is introduced into the music recommendation system,and the Relation model is applied to music recommendation.Use a self-encoder to encode user features and music audio features,and extract their high-level features.In the music encoding process,the user characteristics and audio characteristics are input to the Multi-Head Attention module,and the two characteristics are combined to achieve personalized music encoding in the encoding stage.Then,combine the user's features of the scored song and the target unrated song,and calculate the similarity between the interest similarity vector y and the feature of the scored song and the target unrated song feature in the Relation Module Degree x,after averaging the two ratings,the user's interest in music between[0,1]is obtained.Evaluation using precision and recall rate in the experiment to verify the validity of the model in the direction of the music recommendation.(4)Finally,a recommendation system based on user behavior and music audio is implemented.The backend of the system uses the Python language to process the files generated by the two methods of training to obtain a recommendation list.The front end uses Java Web related technologies to implement user behavior recording modules,recommendation modules,etc.,and the recommendation results generated by the two methods are displayed through a WeChat applet.