Collaborative Filtering Recommendation Algorithms Based On Similar User Search

With the rapid development of Internet technologies,we can not only get a large amount of information efficiently from the Internet,but also constantly upload new information to the Internet,which leads to the explosive growth of information on the Internet.To solve the issue of information overload,recommendation system algorithms have been proposed.The core idea behind recommendation system is to explore users’ personalized preferences based on users’ historical behavior data,and recommend items that match users’ interests to users.In recommendation system,the concept of collaborative filtering is widely used because of its strong interpretability and domain independence.Generally,collaborative filtering algorithms consist of two stages:similar user search and rating prediction.At present,issues of data sparsity and rating diversity still affect the recommendation performance of collaborative filtering algorithms.In order to alleviate the above issues,this thesis mainly studies the similar user search methods for collaborative filtering recommendation.By introducing the concepts of multiorder nearest neighbor,rating subspace and trust mechanism,this thesis alleviates the issue of data sparsity and rating diversity well.The proposed algorithms have good prediction performance and improve the recommendation accuracy.The innovative work of this thesis is summarized as follows:(1)This thesis proposes a multi-order nearest neighbor(MONN)method for collaborative filtering recommendation.In collaborative filtering algorithms,users with high similarity are always regarded as similar users.However,the measurement of similarity between users may be not so accurate owing to sparse rating data.To improve the quality of similar user set,this thesis defines the concept of multi-order nearest neighbors and proposes MONN.On the basis of multi-order nearest neighbors,MONN algorithm designs a new similar user search method and an iterative rating prediction method.Experimental results on four real datasets verify that MONN can not only generate a high-quality similar user set,but also improve the recommendation performance.(2)This thesis presents a user rating subspace(URS)method for collaborative filtering recommendation.MONN directly deals with the rating data and does not consider the rating diversity that has an effect on the search result.To further improve the quality of similar user set,this thesis puts forward URS.First,URS partitions the rating range into three intervals:high,medium and low.On the basis of these three intervals,the set of items for a user is divided into three user rating subspaces constructed with high,medium and low ratings.According to the user rating subspaces,this thesis designs a new similarity measurement that is applied to similar user search and rating prediction.Experimental results on real data show that URS can alleviate the issue caused by rating diversity and achieve better prediction performance.(3)This thesis proposes a user rating subspace and trust mechanism(URSTM)method for collaborative filtering recommendation.To improve the reliability of similar user set,this thesis introduces trust mechanism and proposes URSTM.The trust relationship used in URSTM comes from the relational data explicitly declared by users,which can supplement the sparse rating data.By using the trust relationship of users,this thesis designs a similarity measurement based on user rating subspaces and trust mechanism and integrates it into the multi-order nearest neighbor search method and iterative rating prediction method.Experimental results show that URSTM can improve the accuracy of recommendation performance and has a better prediction ability.