Research On Personalized Recommendation System Based On Item Clustering And User Interest

With the rapid development of the Internet and spread of e-commerce, the scale ofinformation and other resources on Internet has got a speed growth. However, a vastamount of information have a serious impact on users to find useful information. In orderto solve this problem, a personalized recommendation system was developed. Theexpansion of the recommender system scale makes the number of users and itemsincreased. Recommendation system is facing a serious problem of data sparsity, resultingin a poor quality of recommendation system. This paper makes a comprehensive analysisof domestic and foreign research, and conducts an in-depth study aiming to solve theproblems of data sparsity and users’ interest change with time in recommendation system.Firstly, we proposed a rating matrix pre-fill method based on clustering techniques.Gathering the items in groups with rating similarity and attribute similarity through theproposed method. In each group, establishing relationships among items with implicitrelationships between items and the attributes of items to satisfy the requirements thatelements in group have attributes similarity and ratings similarity attributes and ratings.In generated items group, compute the similarities of items and select n items as thetarget item’s neighbors, predicting the rating value of items have no rating and pre-fill thematrix;Secondly, a collaborative filtering recommendation algorithm based on changes ofuser interests is proposed. Because users’ interests will change over time, we suppose thatthe nearly rating time should be assigned a higher weight and the rating time is closer thehigher of users similarity. So, using a time utility function to assign different rating withdifferent utility value during the calculation of similarity. Thereby improve the accuracyof similar neighbors and the quality of the recommended result.Finally, we compare the experimental evaluations and analysis of the algorithmsproposed in this paper with the traditional methods and other existed algorithms on datasets with different sparsity.