Collaborative Filtering Method Using In Recommender System Of E-commerce

Along with leaping development of network and information technology services, the mass electronic data of commodities and services in e-commerce make lots of users lost in seas and oceans of electronic message when providing more information service for users, finding messages user wanted is difficult. The recommender system which is part of intelligent service has become an important and irreplaceable technology in personalized recommendation services in e-commerce. Therefore, studying the recommender module in e-commerce has good prospects and practical applications.Based on personalized recommender system that extensively used in the field of e-commerce at present, we introduces different recommender technologies in e-commerce and studies advantages and disadvantages of them respectively, paying more attentions to the collaborative filtering (CF) technology, specially for studying the filling modules in CF, finally, propose an collaborative filtering method base on filling modules for addressing disadvantages of CF such as data sparsity, accuracy of similarities and use the method in recommender modules in e-commerce.On the research of filling modules, this article indicated two alternative filling strategies. First, this strategy use native Bayesian method to make statistical analysis on attributes of user and item and building a characteristic attribute table for predicting ratings which is on behalf of Bayesian model, this attribute table combines with the item model composed by item similarities to establish a weighted way for filling the user-item sparsity matrix. Another strategy is based on user clustering and item clustering, through using user group-item rating matrix and user-item group rating matrix respectively by user clustering and item clustering to fill the user-item sparsity matrix. The experiment results show that collaborative filtering methods based on filling modules have higher performance than traditional collaborative filtering, especially on accuracy of similarities about user and item and improving accuracy of recommender system's recommendations. At the same time, because different recommender modules are building from different data sources, avoiding the recommender system excessively depend on one data source; effectively combining the contend-based method and rating-based method together in recommender modules, the recommender system has more strong robustness and performance.