Non-negative Matric Factorization And Clustering Methods Applied Research In Personalized Recommendation System

With the popularity of the Internet, E-commerce has become the mainstream currently. Alarge amount of information on the Internet is growing exponentially, leading to theemergence of “information overload” phenomenon. A99%of information may be redundantfor our customers. It is very difficult for users to quickly find the products of their interests insuch a flood of information on Internet. Therefore, the personalized recommendation systemin E-commence has been emerged recently. Among all the techniques in E-commence,collaborative filtering recommended technique attracts more attention in both commercebusiness communities and researchers, which is also a hot topic within E-commence.However, in most practical application, there are still many problems existed, such as datasparsity issues of user’s evaluation to items and the lower evaluation score for all the items. Inaddition, the performance of recommended system will be degraded along with the increaseof users and items. Based on these current issues, this paper proposes a new strategy toimprove the traditional collaborative filtering algorithm.The following are the major work completed in this research:(1). The domestic and foreign research progress for the current recommendationtechniques has been systematically elaborated. Moreover, the classification, the structure andtechnical issues of the recommendation system have been analyzed and discussed in detail.(2). The potential issues such as the data sparse in collaborative filteringrecommendation algorithm have been addressed. The characteristic of user-basedcollaborative filtering and item-based collaborative filtering techniques has also discussed inthis paper. A non-negative matrix factorization technique is proposed to deal with the highdimensional data problems, in order to alleviate the data sparseness problem with promisingresults.(3). To solve the data sparseness and scalability problems in traditional collaborativefiltering technique, a novel strategy has been proposed. First, applying the non-negativematrix factorization technique to regulate the user and item score matrix, then simplify andlower the dimension. Second, clustering algorithm is employed to divide the various usertypes, and the dividing results based on clustering algorithm are used as nearest neighbor.Lastly, the collaborative filtering algorithm is utilized to predict the score and makerecommendation.(4). The proposed methodology was verified through comparative experiments based onthe data from MovieLens website.