Research On Algorithm For Citation Of Hybrid Music Recommender System

Since the establishment of the International Society for Music Information Retrieval in 2000,there have been numerous researches focus on music recommendation system.A significant number of recommender systems and methods have been set up.However,the challenges for the music recommender system are still existing,such as the feedback data made by consumers sparsity problem,and the songs data with long tail distribution,etc,which are chronic problems.In this thesis,the algorithm for hybrid music recommendation are researched.We organize the recent research results of hybrid music recommendation,and then a deep neural network hybrid recommendation(DNNHR)are presented to solve the above problems.We experiment with the recommendation’s performance,accuracy,coverage and novelty.Meanwhile the collaborative filtering recommendation’s,content-based recommendation’s and neural network collaborative filtering recommendation’s evaluation results are presented.The properties of these recommendations are analyzed through comparing the evaluation results.Finally,we experiment with the hyper-parameters of DNNHR,and then we make suggestions about selecting values to hyper-parameters.The research shows that:⑴The implicit feedback data’s sparsity can be reduced through setting screening threshold for the features such as the average number of plays.⑵TF-IDF method and principal component method are used to process the text data,which can effectively utilize the song information as well as preventing the feature dimension from being too high.The evaluation experiment shows that:⑴The training speed differences between the four recommendations are not obvious,and the time difference of 10~5 iterations is about 1 second.⑵The recommendations which combine with the neural network make an accuracy improvement.⑶The collaborative filtering recommendation’s coverage metric and novelty metric are lowest among all recommendations in experiment.The hyper-parameter experiment shows that:⑴The grater number of layers,learning rate and decay rate in neural network are not always the better,which means we should find an optimal selecting value range.For the music recommendation,the neural network hidden layer to take 3 layers,learning rate to take 0.001,the attenuation rate to take 0.001 is appropriate.Based on the above results,the thesis concludes that:the DNNHR can settle consumers sparsity problem and process the songs data with long tail distribution.In addition,the DNNHR can make an accuracy improvement based on other three recommendations.