Research On Collaborative Filtering Based On Recommender Systems

With the arrival of the Internet, E-commerce has become the mainstream currently, As thenumber of user transaction volume increases rapidly, the data of network growths as the trendof exponential, which led directly to the "information overload" phenomenon intensified. In thevast amounts of network data information, the user should be difficult to choose their owninterest information. Therefore, the E-commerce platform in the personalized recommendationsystem came into being.In the1990s, personalized recommendation technology was as an independent concept.After that, a variety of recommendation techniques are rapidly develop in recent years. Forexample: information retrieval, content-based recommendation, association rules-basedrecommendation, collaborative filtering technology as well as mixed recommendation. Currently,the most widely used personalized recommendation technology is collaborative filteringalgorithm. Collaborative filtering technology is the best one of personality recommendationalgorithm. With the number of users and commodities of the e-commerce platform increasingsharply, there are some problems of collaborative filtering such as cold start, the sparse ofUser-Item rate matrix, the scalability of system to be solved.Collaborative filtering is divided into memory-based collaborative filtering algorithm andmodel-based collaborative filtering algorithms, memory-based collaborative filtering algorithmis mainly applied K-nearest neighbors(KNN) algorithm including the user-based and item-basedapproach and the predominant algorithm to model-based collaborative filtering algorithm ismatrix factorization(MF) including singular value decomposition(SVD) and latent factormodel(LFM). However the user’s interests in different time haven’t been taken into equalconsideration with the KNN method being used, which leads to the lack of effectiveness in thegiven period of time. To solve this existing problem, this paper presents an improvedcollaborative filtering algorithm, which based on Shrinkage and time weight, to make one ratingmore approaching the other user’s rating the greater weight of recommendation it process, andaccording to the time weight to enhance the acquisition of similar user, thereby to improve theaccuracy of the recommendation; Singular value decomposition (SVD) was first used in matrixfactorization collaborative filtering, Combining the singular value decomposition and the nearestneighbor algorithm can improve the accuracy of collaborative filtering algorithms, but the timecomplexity of singular value decomposition is very high, scalability is not strong, this paperpresents combine the latent factor model and the K-nearest neighbors algorithm to reducing thetime complexity while improving the scalability.