Research On Collaborative Filtering Algorithm In Recommender Systems

The modern age is information overloaded with a rapid developed Internet technology.To solve the problem of information overload,many researchers begin to study recommender systems in order to handle the big data.Recommender systems mine data in the history of users' behaviors,model users' interests,and then predict users' future behaviors.They connect users and information.Recommender systems not only help users to find out useful information,but also provide the visualization of such information to users,so they can lead both information consumers and producers into a win-win situation.There are two main problems in recommender systems:one is rating prediction problem,the other is Top N recommendation.Collaborative filtering is the most widely used and most successful recommender system technology.This paper mainly dives in solving the rating prediction problem in collaborative filtering situation,which aims to increase the accuracy of the rating prediction.In general,the main contents of this paper include the following aspects:First,this paper proposes an improved calculation method of similarity based on the memory-based collaborative filtering recommendation.Calculation of the similarity is crucial in memory-based collaborative filtering,because it affects both the choice of neighbors and the weight of rating prediction.When the rating matrix is sparse,the number of common ratings by two users can be very small while the ratings are very close to each other,which will produce an incorrect higher similarity.To solve this problem,this paper proposes an improved similarity measure which takes both the value of ratings and the number of common ratings into consideration.What's more,this paper offers a method that considers the implicit feedback,and finally presents a new similarity.Experimental results show that the improved method can increase the prediction accuracy fairly well.Second,this paper proposes an improved memory-based collaborative filtering approach based on the matrix factorization collaborative filtering method.In the matrix factorization model,the rating prediction problem can be transformed into the user-item rating matrix completion problem.In the rating prediction formula of memory-based approach,the rating prediction benchmark is the average score of the users or items;while in the matrix factorization approach,the rating prediction benchmark can be divided into three parts and is more closer to the real situation.Therefore,in this paper,we use the rating prediction benchmark derived from matrix factorization collaborative filtering method to replace the benchmark in rating prediction formula of memory-based approach.Experimental results show that the improved method can effectively increase the prediction accuracy.