A Music Recommendation Algorithm Based On User Preference Auto-Annotation Using CNN

With the rapid development of web service in recent years,users can easily use terminals like smart phones to listen to music online.Because of the huge amount of musical data,users have to face the overload of information.Therefore,it is necessary to offer a personalized music recommendation service for the users.Traditional recommendation systems based on labels need a lot of manual work to annotate music and maintain label models and the resulting multidimensional labels can easily make the recommendation model complicated;while recommendation systems based on neighborhood need to collect large user data and have relatively high computational complexity.Therefore,in the recommendation system it is difficult to recommend music to the right users when new music without user data or labels is added to the library.Regarding big data,compared with the method based on collaborative filter,the music vector model based on word2 vec can describe the similarity between songs more precisely.And the convolutional neural network has advantages in extracting features automatically and it can effectively extract the musical features from mel-spectrogram.Therefore,we obtain the user preference based on word2 vec music vector model,classify the music according to user preference using the convolutional neural network,and annotate the user preference and make recommendation in the case of cold start.Our contributions are as below:First,targeting at the problems of manual labeling,we propose a user preference clustering method based on word2 vec music vector model.This method reduces the dimension of music vectors and clusters them to obtain the user preference classification of music.Since the music in the user collection can be classified more centrally according to user preference,the proposed method can effectively describe user preference and make a better recommendation.In the meantime,this method can also offer user preference labels for the training samples of all kinds of recommendation algorithms.Secondly,targeting at the cold start scenario when new music without user data and labels is added to the library,we propose a music recommendation method based on user preference auto-annotation using convolutional neural network.Compared to existing algorithms,the proposed method can annotate unknown songs according to user preference automatically in the cold start scenario and effectively recommend the music to the corresponding user group with the same preference.