The Research On Collaborative Filtering Algorithm Of Personalized Recommender

With the rapid development of network resources and E-commerce, Internet users are also on theincrease. With dazzling resources on the Internet, Personalized recommendation comes into being in orderto help Internet users find out the resources they need. It transforms from previous commodity needed tofind by oneself to system actively pushing commodity in which users are probably interested, which bettersolves the problem of information overload. So it can keep regular customers, attract potential customersand make profit for companies. Personalized recommendation recommends different commodity todifferent users, which meet people’s need of personalization. As the most successfully and widely usedrecommendation technology in personalized recommender systems, Collaborative filtering has been widelyapplied to various commercial websites and e-library. However, some bottleneck problems exist inCollaborative filtering, such as scalability, sparsity, cold-start and security.Collaborative Filtering is divided into memory-based collaborative filtering and model-basedcollaborative filtering, based on the theory of model-based collaborative filtering, the process of Bayesianmodel fill and user cluster have to be carried on off-line, online recommendation. It solves scalability andcold start problems. Based on the theory of memory-based collaborative filtering, this paper proposes anew collaborative filtering algorithm based on users local preference similarity to solve sparsity problems.Research works are taken as following:(1)To solve the scalability problem,this paper proposes a user cluster collaborative filtering algorithmbased on the Bayesian model fill. In the process of predicting ratings that users have not rated in the unionof user rating items by Bayesian model, not only the user’s attribute, but also the item’s attribute should beconsidered. The user weighting factor is introduced into rating prediction and improves the selection ofinitial cluster centers. Considering the scalability, the process of Bayesian model fill and users’ cluster haveto be carried on off-line. The algorithm solves the cold-start and scalability problems. The experimentalresults show that he algorithm efficiently improves the recommendation quality and is relatively suitable tothe higher scalability field.(2) To solve traditional similarity measure methods work poor in the extreme sparsity of user ratingdata,this paper proposes a new collaborative filtering algorithm based on users local preference similarity. This paper use localized preference to calculate similarity of the users, under this condition that the nearestneighbor users rate little about the target users’ not rated items, combing the users’ interest and slope-onealgorithm to predict item ratings, better solves the problem of sparsity and proves this algorithm’sfeasibility by the experiments. This algorithm is relatively suitable to the field in which system is initiallyset up, the data volume is not very big and the accuracy of recommendation is strong.