A Research On Similarity Calculation In Recommender Systems

According to statistics, the proportion of unstructured data has reached about 80%.The rapid growth of Internet scale and coverage led to the famous problem called "information overload". On the one hand, abundant information resources provides people with great help; On the other hand, users can’t get useful knowledge from so much information, because of this, the information efficiency is greatly reduced.In order to solve this problem, people need an information filtering method. Through this method, people can get resources that is useful to themselves. Recommender system is important means of information filtering and effective way to solve the information overload problem. In E-commerce recommender system, social networking sites and information platforms, collaborative filtering technology is the most popular and successful method at present.The recommender system based on collaborative filtering is the most widely used and further studied.The key of this algorithm is to find user neighbors or item neighbors, and the accuracy determines the quality of the final results of the recommendation. Because of the neighbors’ finding relies on similarity calculation of users and items,the design of an appropriate calculation of similarity is the key issue of a successful recommendation algorithm.This paper firstly introduces three kinds of common recommendation method and several evaluation indicators; Then analyzes collaborative filtering algorithm by experiment. According to experimental result, this paper summarizes the performance of recommender system under different number of neighbors.After that,the paper also introduces several similarity calculation methods and their applicable scenes, the advantages and disadvantages.In order to improve the accuracy of similarity calculation, this paper discusses four factors that may affect the accuracy of the similarity calculation, including high frequency items, the number of common grade users, the grade weight of item and the similarity weight. Improved algorithms for these factors are given out.And then offline experiment proves that the performance of recommender system based on improved similarity calculation is superior to the one based on traditional similarity calculation.At last, the paper simulated a recommender system based on collaborative filtering algorithm, SlopeOne algorithm and SVD algorithm by Taste tool in Mahout.