Research On Collaborative Filtering Algorithm Of Personalized Recommendation

With the rapid development of network technology and information service,the application of Internet service is becoming mature,such as social network and e-commerce website.Personalized recommendation plays an important role in social and economical network service,as it is a kind of key technological factor.Collaborative filtering is broadly applied in personalized recommendation because of its excellent modeling technique and intelligent service.It accelerates the development of personalized recommendation.However,under the situation of data sparsity,collaborative filtering is facing a series of problems: the inaccurate similarity measurement and the declining of recommendation precision.To solve these problems,this article made a research on collaborative filtering of personalized recommendation.The main works are as follows:1)Aiming at overcoming the flaw of traditional similarity measurement,user rating behavior based similarity model(URBSM)is built.User behavior patterns and the overall data environment changes are mastered in the aspect of user behavior by URBSM.Based on the similarities and differences of user behavior,nonlinear S function is introduced into the measurement model.By considering the overlap of user rating items and the rating habits,we double weight the similarity by both Jaccard similarity coefficient and mean square distance similarity MSD,which can enhance the effect of user's interest's difference.2)Faced with the poor recommendation precision that caused by rare data of neighbor,grey rating prediction model(GRPM)is built.According to the characteristics of “poor information” and “small sample” of collaborative filtering,combine grey prediction with collaborative filtering technology.So grey rating prediction mechanism is established.The problem that the data foundation is rare during rating prediction which is caused by few data in neighbor is solved by rating sequence pre-processing,and it is put forward that the average rating value of the item is used to fill the vacancies of the initial rating sequence.Increment rating sequence is formed according to the value of similarity in grey prediction.And the average value of the item's ratings is used to fill the initial rating sequence which has vacancy,which makes rating closer to reality.And accumulation operation is used to strengthen theinfluence and impact on rating prediction from the user whose similarity is bigger.3)Compare URBSM with traditional similarity calculation by experiments.And compare GRPM with traditional collaborative filtering prediction method.Conduct contrast experiments to verify the efficiency of algorithms from three aspects: different neighbor scale,different recommendation list and different sparse level.Finally,form improved collaborative filtering by combining URBSM with GRPM,and compares its experimental results with traditional collaborative filtering,PM-CF and SVD collaborative technique.