A Clustering-based Collaborative Filtering Recommendation Algorithm

With the development of Internet and network technology,people are surrounded by a large quantity of information and data.Information overloading,which has been becoming a problem,gives rise to many demands both for persons and enterprises.To solve this problem,Recommender System(RS)emerged.Recommender algorithms are the key realization process of RS.Collaborative Filtering(CF)recommender algorithm,which is one of the recommender methods with characteristic of simple and straightforward,is mainly based on user history rating which is called utility matrix.Memory-based Collaborative filtering has two major process:neighbor similarity measurement and predictive scoring.Clustering,one of the unsupervised learning methods,can be used in the process of neighbor similarity measurement to help CF to get high efficiency and effectiveness.Improved Clustering by Fast Search and Find of Density Peaks with Data Field(IMP-CFSFDP-DF)is utilized to model users before CF process.In addition,nine Combinations of CF process are used to explore the laws of CF results influenced by them.Firstly,it is found that multilevel high order difference can locate the position of cluster centers in CFSFDP-DF and then separate them from other points.It improves the problem of manual determination of cluster centers only by artificial observation in CFSFDP-DF The validity and reliability of the novel approach we proposed are demonstrated by experiments of various test cases.The algorithm we proposed is called Improved Clustering by Fast Search and Find of Density Peaks with Data Field(IMP-CFSFDP-DF).Secondly,IMP-CFSFDP-DF,CFSFDP-DF and K-means methods are compared with each other.It is found that IMP-CFSFDP-DF can automatically and accurately find different types of regions in two dimensional data sets quickly,which shows a better performance than CFSFDP-DF and K-means.Thirdly,a collaborative filtering recommendation with IMP-CFSFDP-DF methods(IMP-CFSFDP-DF CF)is constructed.IMP-CFSFDP-DF method is applied to model user characteristics.Thus,IMP-CFSFDP-DF methods is extended to three dimensions and the comparison times are reduced.Fourthly,the validity and efficiency of IMP-CFSFDP-DF CF is demonstrated by experiments.Using neighbor similarity comparison and predictive scoring process,a combination method with nine combinations of CF with IMP-CFSFDP-DF process is tested.Utilizing these methods above with two sets of data extracting randomly from publicly datasets,a good performance of recommender can be reached.Compared to CF with k-means,these methods can achieve both low Mean-Absolute-Error(MAE)and Rooted Mean Square Error(RSME)speedily with less similarity comparisons.Furthermore,we give some suggestions about the choices of combination of similarity measures and scoring methods.In addition,it is found that clustering preprocessing is applicable to recommenders needing high efficiency while not having the ability to stand too many neighbors comparisons.