A Music Recommendation System Based On Multi-Modal Fusion

In view of the increasing richness of music forms,the underlying features of music are often overlooked when using traditional collaborative filtering.A music recommendation system based on multi-model is proposed by fusing audio features and lyric information of music and correlating the fused feature information with users' preferences.This paper mainly discusses the method of audio feature extraction.MFCC is chosen as the standard of audio feature extraction,while Doc2 Vec is used to process the lyric features of text attributes.To solve the problem of multi-modal fusion,an EFFC multi-modal fusion method based on HMM constraints is proposed.After collecting user ratings,mapping and similarity analysis between multi-modal features and user ratings are carried out to find out some songs that users are most interested in and recommend them.At the same time,on the basis of multi-modal music recommendation system,the problem of cold start in recommendation system is solved by real-time heat recommendation.In the experiment,the optimal recommended number of songs was determined by plotting the PR curve.In view of the system's ability to measure accuracy and weaken the long tail effect of the recommendation system,the accuracy,recall and coverage rates are adopted respectively,and three kinds of music recommendation systems are compared.The results show that compared with the traditional single-model music recommendation system,the accuracy of the system decreases by only 3%,while the coverage increases by 28%.This proves that the system can greatly improve the ability to excavate the long tail with little loss of accuracy.