Collaborative Filtering Algorithms Based On Multi-factors Research Framework

"Shop around" can help customers to get more information,but the situation has changed dramatically in e-commerce nowadays.In the Internet era,the rapid growth of the World Wide Web leads to a serious problem of information overload:customers can not shop around.There are too many different stores and various products to decide to buy in e-commerce system.The real situation is that customers are also puzzled how to get what he really wants.As a powerful tool,recommendation system emerges to help people out of the overloaded information,which therefore attracts great interest of scientists."Multi-factors research framework of collaborative filtering recommendation algorithm" is proposed in this paper.The traditional research framework just consider "behavior commonality"without considering other various influencing factors.This leads to two major problems.Firstly,data sparse problem.There are no common collected items between many users.So we can not analyze their tastes.Similarities between users who have no common collected items can not be calculated.Secondly,considering the single factor only.The traditional research framework does not consider other factors,such as "identity of taste" and "attribute correlation".To cure the above problems,a new research framework named "Multi-factors research framework of collaborative filtering recommendation algorithm" is proposed in this paper.It considers three major factors.It not only consider "behavior commonality",but also consider "identity of taste"and "attribute correlation".Three recommendation algorithms based on collaborative filtering are proposed in this paper,which name are "meta-similarity collaborative filtering","global approval degree collaborative filtering" and "focus symbol degree collaborative filtering".We use two datasets,namely,MovieLens and TMall dataset to judge the performance of algorithms proposed in this paper.Many weighting methods are applied in this paper.The algorithms proposed in this paper are useful for easing data sparse problem,improving precision and personality.Under the new research framework,taking into consideration "identity of taste" and"behavior commonality",collaboritive filtering algorithm based on meta-similarity is proposed in this paper.This idea can be used to improve many traditional algorithms.Specifically,two different meta-similarity algorithms is proposed,the one is based on pierce correlation coefficient and the other one is based on NBI.Numerical results show they perform much better than traditional algorithms.Then a modified collaborative filtering algorithm based on standard meta-similarity is proposed,it take "tri-meta similarity" into consideration.It performs much better than the standard meta-similarity algorithm.Under the new research framework,taking into consideration "attribute correlation" and"behavior commona]ity",collaboritive filtering algorithm based on global approval degree is proposed in this paper.This idea can be used to improve many traditional algorithms.Specifically,two different algorithms is proposed,the one is based on pierce correlation coefficient and the other one is based on NBI.Experimental results show the algorithm works well not only on algorithmic accuracy,but also on personalized recommendation.Under the new research framework,collaboritive filtering algorithm based on focus symbol,degree is proposed.It is different from global approval degree algorithm that it consider more"identity of taste" than "attribute correlation".Experimental results show the algorithm works well not only on algorithmic accuracy,but also on personalized recommendation.Taking into consideration various factors,a new research framework is proposed in this paper.The algorithms based on the new framework perform much better than traditional algorithms.