The Research On Collaborative Filtering Techniques In Personalized Recommendations System

Personalized recommendation system is the inevitable outcome of the Internet age aswell as the development of human civilization. In face of a complex variety of Internetinformation, people tend to feel overwhelmed and cannot find the things they need, so therecommendation system was born in this demand environment. Recommendation system isa kind of intelligent systems which can provide satisfactory information for peopleautomatically, while, personalized recommendation system is the kind of intelligentsystems which can provide satisfactory information for people as well as meeting theindividual needs for people.Collaborative filtering technology is the most successful technology of thepersonalized recommendation system, it took advantage of User-Item rating data, In acollaborative way between users or items，then, generate recommendations. However，asinformation continues to expand, collaborative filtering technology still has a lot ofproblems need to be studied and solved. Among them, the rating data sparsity and the coldstart problem and other problems are the main factors that hinder the further developmentof the accuracy and efficiency of recommendation system. This paper researchs from thedata sparsity and cold start problems, from the accuracy point of view, it mainly includesthe following two aspects.(1) For the data sparseness problem, it proposes a improved collaborative filteringalgorithm based on k-means item clustering. As the traditional collaborative filteringalgorithm based on k-means item clustering ignores the inaccuracies problem with datasparsity calculation between items, which leads to the poor recommended accuracy. Thispaper presents improved methods. Firstly, by data preprocessing process it ensure dataintegrity in items clustering. Secondly, it divides the item clusters which combines withK-means clustering algorithm. And then, in the nearest neighbor query process, accordingto the number of items in the cluster which the target item belong to, it chooses thatwhether need to carry out the similarity calculation again for the items, and then selects thecollection of nearest neighbors. Finally, it predicts rating according to the similarityweighted scoring formula, and generates recommendations ultimately. (2)For the new user cold start problem, it proposes the concept of informationextension to solve this problem. By global-item information extension techniques, it solvedthe problem with low accuracy because of minimal score. Firstly, it need to get a collectionof similar items of the target item, and extended scoring information about the new user byscoring datas from similar items, and then it gets the target user’s nearest neighborsconnection with the calculation of similarity between users. Finally, according to thesimilarity weighted scoring formula to predicted rating it generates recommendations. Inthe paper, it uses the information extension techniques for the new user with the user-basedcollaborative filtering algorithm, and can raise the new user similarity computability,thereby enhancing the accuracy of the recommendation algorithm.