Deep Learning Based Music Recommendation System

With the rapid growth of online information,there has been a serious problem of information overload.Simple information retrieval systems can no longer meet the needs of users.Therefore,Recommendation system came into being.In recent years,deep learning has achieved great success in many fields.However,there are few applications and researches on the recommendation system.This thesis focuses on music recommendation and proposes a music recommendation system based on deep learning.The research steps are as follows.Firstly,a universal recommendation scoring function is defined based on the MSD data set.Secondly,the MFCC audio features and Chroma audio features are extracted and lyrics features are obtained through the word embedding method.Thirdly,a hybrid recommendation model based on convolutional autoencoder was built.Fourthly,we evaluated the quality of the recommendation.Finally,A front-end recommendation platform was built.The points of this thesis are as follows.The hybrid recommendation model combines the content features and collaborative filtering.It uses two Autoencoders to learn the user features and the song features respectively.Song-side Autoencoder uses a convolutional layer to learn the audio features and a full connection layer to learn lyrics features.User-side Autoencoder uses a full connection layer to learn the rate vector.After the pre-training,a tightly coupled model is trained.Kaggle's million song dataset challenge provides 386213 songs.The validation and test sets combined contain 110k users,half of their history released.The baseline is set as a memory-based collaborative filtering algorithm.Then,A matrix factorization method and the hybrid recommendation model are implemented.The MAP@500 of the hybrid recommendation model can reach 0.23682,which can be improved by 0.05 compared with the baseline.The experimental results show that the performance of the model compared to traditional collaborative filtering method is improved significantly.