Design And Implementation Of Music Recommentdation System Based On Original Tags

With the vigorous development of major music websites,using music apps has become an indispensable part of people's lives.In order to better satisfy users' search for favorite music songs,it is more and more important to use algorithms to build a music recommendation system for song distribution.Among them,the collaborative filtering algorithm is widely used in the recall process of major music recommendation systems.The traditional collaborative filtering algorithm has a coarser granularity in characterizing the similarity between users and music,and the recommendation recall results are all concentrated in the head popular,and the degree of personalization is weak.Therefore,how to use mining user information and music song information to improve the traditional collaborative filtering algorithm,carry out more effective recommendation recall,and improve the overall recommendation system effect.It is the main research direction of many scholars.This thesis combines the company's actual K song App music recommendation system project to study the music recommendation system recall algorithm.The main research content includes the following four points:(1)Aiming at the problem that the original tag data is small and the classification based on the LDA model is not effective,a LDA music theme representation method based on lyrics keywords supplementing the original tags is proposed.This method uses the TF-IDF algorithm to extract lyrics keywords to supplement the original tags to improve the integrity of the text,and then trains the music theme classification through the LDA model.Experiments show that this method can improve the learning effect of the LDA model and achieve the purpose of accurately representing the music theme.(2)Aiming at the problem that user interest is difficult to be accurately portrayed,a user interest expression method based on graph embedding technology is proposed.This method digs out user interest preferences from user behaviors,and uses the Node2vec model to model the music song of the user's positive feedback behavior in the session,and learn the corresponding music song vector.Then,different weights are defined for different behaviors of users,and the vector sequences of music and songs that are positively fed back by users are weighted and merged to obtain a vector representation of user interest.Experiments show that the accuracy of expressing user interest through this method is high,and the user interest is more detailed.The vector obtained by the fusion can well represent the user's interest.(3)Aiming at the problem that the original collaborative filtering algorithm focuses on recalling songs that are concentrated in popular songs and the degree of personalization is not enough,an improved collaborative filtering algorithm that combines music theme information and user interest information is proposed.The music theme information obtained in the previous part is used to calculate the similarity of the music theme,and the theme similarity is added as a weight to the music similarity calculation of the ItemCF algorithm.According to the obtained user interest information,the user interest similarity is calculated,and the obtained similarity is added as a weight to the user similarity calculation of the UserCF algorithm.Experiments show that after adding music theme similarity and user interest similarity,the improved collaborative filtering algorithm is more effective than the traditional collaborative filtering algorithm,and the recall effect is significantly improved.(4)On the basis of the above three contents,through the integration of the music theme presentation layer,the user interest presentation layer,the music song recall recommendation layer,the original fine sorting and re-ranking layer and the AB test framework are used to achieve a complete basis Original label music recommendation system.