Font Size: a A A

Research On Some Key Issues Of Recommender Systems

Posted on:2013-01-08Degree:DoctorType:Dissertation
Country:ChinaCandidate:L RenFull Text:PDF
GTID:1118330374968011Subject:Systems analysis and integration
Abstract/Summary:PDF Full Text Request
With the rapid development of Internet, the application of IT has been extended from professional domains to human life. Various Internet information services and applications have entered into human society, and human physical activities are evolving to online virtual ones. With the popularity of such applications as SNS, cloud computing and P2P, the working form of Internet application are changing from one-to-many to many-to-many, and users are both the customer and producer of information. With the explosive growth of online information, it is hard for users to filter the information in large information space according to their information needs, causing the issue of information overload. In academia and industry, a lot of research and practice have been done on the issue of information overload, and lots of personalized solutions have been proposed as well to provide users with information on demand.As an intelligent personalized information system, recommender systems describe the user's lone-term information need by user modeling, based on which it can customize the personalized information with the specific recommendation strategy. For its advantage of user-driven mechanism, active service and personalization, recommender systems have been widely applied in the fields of E-commerce, online learning and digital library, and taken as the most promising direction in information personalization. Some achievements have made for the research and application of recommender systems, yet a few issues existing in recommendation approach have emerged with the increasing scale of user and information, in which the issues of sparsity and concept drift are the major factors hindering the further development and application of recommender systems.To follow the development of Internet information service and improve the accuracy of recommendation, this work took it as the object to alleviate the influence of sparsity and concept drift, and discussed the following subject:(1) The status of Internet information service and recommender system was summarized. By reviewing the history of recommender systems, the issues in recommender systems were elaborated.(2) The main cause of information overload was analyzed. Based on the comparison between information retrieval and information filtering, the research on recommender systems was made in the aspects of user modeling, classification of recommendation algorithm and evaluation.(3) With analyzing the cause and influence of matrix sparsity, its forms of direct and indirect impacts were summarized. Incorporated with the existing solutions for sparsity, this work t summarized he main methods and strategies in resolving the issue of sparsity.(4) Aiming at the impact of sparsity on the computation of similarity, the excessive dependence on public ratings of item-based similarity was discussed. For the reason that the issue of sparsity reduce the public ratings, the similarity computing based on few ratings will cause the decrease of accuracy correspondingly. For the above reason, a weighted similarity-boosted collaborative filtering approach (WSBCF) was proposed, in which, by introducing the rating overlap factor in similarity computing, the classical similarity was amended to improve the accuracy of recommendation.(5) The procedure of rating prediction in classical collaborative filtering is an ideal decision process based on non-sparse rating matrix. With the sparse rating matrix, the personalized rating prediction usually results in decreasing accuracy of recommendation. Aiming at the impact of sparsity on rating prediction, and incorporating the user's conformity presented in recommendation, an item-based collaborative filtering approach integrating balanced prediction (IBCFBP) was proposed. To improve the procedure of rating prediction, IBCFBP integrates the personalized rating with the global rating in rating prediction by levering the importance of both ratings based on the distribution of global rating.(6) For the reason that each procedure of collaborative filtering relies on rating matrix to realizing respective function, the sparsity of rating matrix can impact all aspects of collaborative filtering. Aiming at the impact of sparse matrix on collaborative filtering with analyzable items, and incorporating the insensitivity of CBF to sparsity, a hybrid recommendation approach based on rating filling (HRRF) is proposed. HRRF improves the data density of rating matrix by employing CBF to fill unobserved ratings.(7) By analyzing the cause and impact of concept drift, this work summarized the influencing approach of the issue. Based on the discussion of existing solutions for concept drift in machine learning and recommender systems, main strategies resolving the issue of concept drift were summarized.(8) Aiming at the impact of concept drift on collaborative filtering, and incorporating the feature of historical ratings'importance decaying exponentially with time, a temporal collaborative filtering (TIBCF) was proposed. To alleviate the disturbance of concept drift, TIBCF employs temporal weights to improve similarity computing and rating prediction.Through the above discussions, the cause and influence of information overload was demonstrated, and the related issues impacting the development of recommender systems was analyzed extensively. Specially, aiming at issues of sparsity and concept drift in collaborative filtering, through improving each procedure of collaborative filtering, four corresponding refined approaches were proposed, and their effectiveness in improving the accuracy of collaborative filtering is demonstrated by experiments. The main contributions of this work include:(1) The cause and influence of information overload is illustrated in view of Internet information service.(2) The requirement of personalized information service was analyzed, and its definition was formulized.(3) Based on summarizing the structure and principle of recommender systems, this work analyzed its related contents of user modeling, recommendation algorithms and evaluation.(4) Through analyzing the cause and influence of sparsity and concept drift, this work discussed the existing solutions and summarized the main solving strategies to both issues.(5) Aiming at the issue of sparsity, its impact is alleviated by improving similarity computing, rating prediction and rating matrix density respectively. Three corresponding refined collaborative filtering approaches of WSBCF, IBCFBP and HRRF were proposed and their effectiveness in improving the accuracy of collaborative filtering was demonstrated by experiments.(6) Aiming at the issue of concept drift, through employing time weight to improve similarity and rating prediction, the reined approach of TIBCF was proposed and its effectiveness is evaluated.
Keywords/Search Tags:Recommender system, Item-based collaborative filtering, Hybridrecommendation, User modeling, Sparsity, Concept drift
PDF Full Text Request
Related items