Research On Hybrid Recommendation Method Of Music Recommender System

With the development of information technology,people are in a dilemma caused by information overload: For producers,it is not an easy task to make their information widely noticed;for consumers,finding the information which they needed is getting harder and harder.In order to help them get out of this dilemma,recommender system comes into being.The main idea of the algorithm is to use the user's behavior history to predict the user's preferences and discover the items that the user may be interested in.However,inevitably,there are two defects which the decrease of performance caused by the sparse and the cold start problem in traditional recommendation algorithm.In order to solve these problems,incorporating the content information into the recommendation,researchers create a hybrid recommender system.However,the content information often has the characteristics of uneven distribution,complex structure,multi-modality,and sparseness,which brings a huge challenge to the traditional hybrid method.The development of deep learning technology provides a new solution to this problem.Deep learning is an end-to-end,non-linear,deep-level network architecture framework,which can uses multi-source heterogeneous data as input and maps them into the same low dimension space.Content information can be used to mitigate cold start and sparseness problem in traditional recommender systems effectively with the combination of deep learning technology and traditional recommendation algorithms.Although the hybrid of deep learning and traditional recommender models has achieved great process,there two important components in the hybrid model: One is the deep component,which outputs the low-dimensional hidden expressions of the extracted content,the other is the traditional matrix factorization components.The combination of these two components is not particularly close and the deep components remain relatively independent while running,which cannot guarantee the relevance of extracted hidden content features and the hidden features of item.Therefore,this paper proposes a Double-Regularization Collaborative Deep Model(DRCDM),which provides feedbacks in both components and imposes regularization on each other.The feature extraction module of the model uses Stacked Autoencoders(SAE),and the traditional recommender module uses Probabilistic Matrix Factorization(PMF).The model updates the two modules alternately during the training process and imposes regularization on two hidden matrix of item content and rating matrix,in order to obtain a closer correlation and potential relationship between two hidden matrix.The hidden layer features are more effective in helping rebuild the scoring matrix.The introduction of item's content features can solve item's cold start problem and the risk of overfitting during the reconstruction process of rating matrix.We validate the recommended performance of our model on MSD dataset which is the largest public music listening history dataset.Compared with three recommender algorithms,the result shows that the DRCDM model has the best recommended performance under the Recall score and the F1 score.