A Study On Collaborative Filtering Model Based On Depth Learning

With the booming of Internet technologies,the way how humans live has changed a lot.In the Internet age full of all kinds of information and competition,it is fatal import for an Internet company to know how to help users quickly and accurately find their interested products and services.And that is why recommendation system technologies came into being.Collaborative filtering is one of the most popular technologies applied in recommendation systems.Traditional collaborative filtering techniques use only the users' rating matrices on items to make recommendations.But in general,those rating matrices are very sparse,which leads to a serious reduction in the recommendation accuracy of the recommendation systems.And in addition,there is a problem of cold start for new items.To address those issues,the main work of this paper is organized as follows:(1)When applying Restricted Boltzmann Machines(RBM)to collaborative filtering,the recommendation performance has a great correlation with the sparseness of rating matrices: when the rating data is very sparse,the recommendation performance is poor.And former RBM-based recommendation systems use only the rating matrices,which causes the problem of cold start for new items.In order to solve those problem,this paper proposes an RBM cooperative filtering method,Content Supplement Restricted Boltzmann Machines(CS-RBM),which combines the content similarity of items.This method calculate the similarity between items after using word2 vec to learn item-content vector,and then the similarity measurement is added to the RBM prediction score so that the final prediction score takes not only the influence of the implicit factor in the rating matrix but also the influence of the similarity between items into account.Experiments on ml-100 k,ml-1m and Netflix data sets show that our method can achieve better recommendation than the original RBM model.(2)Since CS-RBM simply makes use of the content information of items,it cannot capture the deeper implicit factor inside that content information for the model improvement,and does not take into account the influence of user characteristics on the model.Aiming at these problems,in this paper,we proposes a depth cooperative model,Double-sided Collaborative Deep Learning(DCDL),which combines bidirectional constraints on user characteristics and item characteristics on the basis of depth cooperative model Collaborative Deep Learning(CDL).The model utilizes the Stacked Denoising Autoencoder(SDAE)and Probabilistic Matrix Factorization(PMF)co-training to automatically learn items hidden features and user hidden features from item contents and rating matrices,so it takes into account both the impact of the item characteristics on the recommendation and the impact of the user characteristics on the recommendation.Experiments on citeulike-a,citeulike-t and Netflix data sets show that DCDL can achieve better recommendation performance than collaborative topic regression(CTR)and Collaborative Deep Learning(CDL).