Research On Collaborative Filtering Recommendation Algorithm Of Multi-matrix Based On Item Attributes

Along with the popularization of Internet and the rapid development of information technology and E-commerce,the "information explosion" and "information overload" are becoming more and more serious,how to find resources needed by user efficiently and quickly has become an urgent problem to be solved.As one of the effective ways to solve this problem,the recommendation system has been widely used in many fields.But its core content-recommendation algorithm,especially the collaborative filtering recommendation algorithm is facing severe problems of sparse user ratings,low recommendation accuracy,poor scalability,cold start and so on.Based on the deep research of traditional collaborative filtering recommendation algorithm,after summarizing the shortcomings and deficiencies of previous research results,the paper proposed collaborative filtering recommendation algorithm of multi-matrix based on item attributes aiming at the problems of the high data sparsity and low recommendation precision.The algorithm first divided item attributes into single value and multi values,and deleted the valueless attributes,preserved the worthy ones according to the fact;Secondly,using the mean method or the scale method to construct the user-item attribute value rating matrix,since there are more than one attribute an item,multi-matrix would be constructed;Third,selected part attributes to participate in the prediction stage according to two criterion: data sparsity and data reduction to ensure the accuracy of the similarity and improve the precision and efficiency of recommendation system on the basis of retaining authenticity of users' preference as much as possible.Fourth,based on each user-item attribute value rating matrix,the preference similarity between users is calculated,and the nearest neighbor set is obtained;Fifth,completed the initial prediction of the target user to the target item for each nearest neighbor set on the user-item rating matrix and got a number of primary ratings;Finally,weighted and summed the primary ratings as a comprehensive rating,and recommended the TOP T items with highest comprehensive ratings to the user.The paper's experimental dataset is the extension of Movie Lens10 M dataset from which selected five training datasets randomly,and then contrasted performance between the proposed algorithm and the traditional algorithm following several steps: the accuracy of initial ratings and comprehensive ratings,attribute selection effect on accuracy,the time and space efficiency,and the new user problem.The experimental results show that the proposed algorithm can effectively improve the recommendation accuracy,and ease the user cold start problem to a certain extent.