Research On User Similarity Function Of Recommender Systems

With the rapid expansion of the data size on the Internet, users can accurately findthe information they need quickly. To address the increasingly serious problem ofinformation overload, a variety of solutions have emerged, and the recommendationsystem is one of the best. Recommendation system is a personalized information service,which can well serve as a bridge between users and information resources.Recommendation system describes the users’ needs through building models, and thenrecommends specific information resources to the target users through somerecommended strategies initiatively. Due to its personal and intelligent features,recommendation system has achieved great success in e-commerce, social networkingsites and video sites, and has become the core subsystems of these application platforms.The recommendation system based on collaborative filtering is the most widely usedand further studied. The key of this algorithm is to find users and project neighbors, andthe accuracy determines the quality of the final results of the recommendation. Becauseof the neighbors’ finding relies on similarity calculation of users and projects, the designof an appropriate calculation of similarity is the key issue of a successfulrecommendation algorithm.This paper introduces the concept and model of recommendation system, andanalyzes its architecture, while discussing the model of users in detail. Then this paperintroduces several common recommendation algorithms and their applicable scenes, theadvantages and disadvantages. After that, the paper also introduces the major similaritycalculation methods and their limitations. At last, the commonly used experimental datasets and evaluation are introduced.In traditional user similarity calculation methods, the weight for each item is thesame. It is known by analyzing that the weight for co-high-rated programs among usersis higher than which for low-rated programs. Considering the group relationship, aweighted similarity calculation method of user is been proposed. This solution willsolve the issue that will lead to decrease the accuracy to find a neighbor. By comparingwith the collaborative filtering algorithm based on traditional user similarity calculationmethod, the experiment, made on MovieLens data sets, shows that the collaborativefiltering algorithm considering the item similarity can significantly improve scoreforecast accuracy and the quality of recommendation results.