Research On Music Recommendation Alogorithm Based On Deep Learning

Fast development in recent years,deep learning approach,the convolutional neural network in NLP,image recognition,and so on are widely used in many fields,and in the field of music recommendation due to the difficult to make full use of music audio features,recommended the effect not beautiful,and most of all because now only music recommendation algorithm by user according to feedback or implicit feedback for music recommendation,The feature of music label information is not fully utilized,which makes the accuracy of user preference prediction low and it is difficult to realize personalized recommendation.Therefore,this thesis proposes a recommendation method based on convolutional neural network combining multiple features of music label information.The basic idea of the recommendation algorithm proposed in this thesis is as follows: firstly,the user preference model is constructed by using the cryptic model matrix decomposition method;Then carries on the pretreatment on audio resources in the system,and distilled to represent characteristics of music MEL frequency spectrum diagram,convolution neural network is used to analyse the music audio feature extraction and fusion of music multiple information(singer ID,year of publication,etc),the music invisible feature vector and user hidden feature vector inner product operation,get the user preference score,According to the preference score,the recommendation list is sorted TOPN to achieve personalized recommendation for users.In this thesis,systematic experiments are carried out on the proposed recommendation algorithm.Based on the convolution neural network model,the network model structure used in this experiment is obtained by integrating the user-music data set in MSD with the Embedding layer.Finally,the model is trained and tested,and the error rate,accuracy rate,recall rate and F1 value are taken as the evaluation indexes of recommendation quality.Experimental results show that the recommendation algorithm proposed in this thesis is superior to other models in different evaluation indexes,and has certain feasibility and effectiveness.Compared to other methods,this thesis make full use of the depth of the neural network automatically extracted features strong advantage,obtain higher music features from the audio content,said at the same time using the Embedding layer into a wide variety of music identification information,make the model has better scalability,build up the characteristics of a more complete system,improve the quality of the music to recommend recommend.