Research Of Recommendation Algorithm Based On Real_Value Restricted Boltzmann Machine

Along with the rapid development of the Internet,there are lots of channels creating information,resulting in explosive growth of infomation.How to obtain useful content from massive data has become a major problem.To solve this problem,recommendation system for specific users was proposed.Recommendation system mine the user's potential interest according to the user's historical behavior information and quickly help users get the interested tems they require from the vast amount of information.This paper focuses on the data sparseness problem of recommendation algorithm and using distributed computing framework to solve the system operating efficiency.There are two ways to solve the problem of data sparsity over time.The first is data filling,which is to establish an effective user model with the user's other valid information.The second is to utilize the user's historical rating information directly and pre-process user history score data through matrix decomposition,clustering,machine learning,and so on.This paper combines the obove two ways,fusing the tag to the Conditional Restricted Boltzmann Machines model.We utilize CRBM's powerful ability to fit arbitrary discrete distribution to predict the missing scores for the user unevaluated products.Specifically,we propose the Real-valued Conditional Restricted Boltzmann Machine model whose visible units are real value firstly.Then,the TF-IDF algorithm of text classification is used to predict the attitude of the user who applied the tags,which multiply with the tag-genome to get the user's scores for products,which are integrated into the user history rating data.In the conditional layer,we incorporate user tagged/untagged {0,1} vector in the original rated/unrated {0,1} vector.we use Spark distributed computing framework to parallelize the algorithm.We not only improve the time efficiency of the algorithm,but also apply more user historical data to improve the accuracy of recommendation.We use real dataset MovieLens to test the algorithm and its parallelization.We use RMSE and MAE as evaluating indicator and 10-fold crossValidation as validation way.The experiment results of Model predicting and Spark parallelization shows that the algorithm proposed in this paper improve the accuracy of recommendation and premote the efficiency.The algorithm hold application value practically.