The Research On Neural_Networks_based User Modeling And Web Information Filtering

The Web personalized service is one of hotspots in AI and information technology. The current information retrieval system which mainly bases on search engine don't concern enough about users' different interests. Users get the same results if they submmit the same query words. At the same time, the good and bad are intermingled. Users have to find suitable information from huge amount of Web pages manually. It's important to improve current information retrieval system with information filtering. Information filtering is the basis of personalized information service.User modeling can enhance the performance of IR. User model is a description about a user group or an individual user. With the user model, computer can acquire, store and restore user's fuzzy dynamic interests. The information stored in user model make up the condition of IR and make IR more effective.In this paper, the current research and application on Internet personalized information retrieval is analyzed. And then Soft Computing including Fuzzy Logic and Nerual Networks are introduced too. According to ANFIS, an improved Nerual Fuzzy networks is introduced into user modeling and web information filtering to satisfy the user. The following key problems are disscussed in this paper.(1) How to express the Web page's content and user's interests. The Vector Space Model is used to map the Web page into a vector Pj. Before filtering, the query words and page examples input by the user is analyzed and mapped into vectors ui too.(2) How to select the character terms to decrease the number of dimensions. In the IF, the object of filtering is retrivaled Web pages. These pages can be divided into twoclasses: one is relevant pages R, another is unrelevant pages r. Taking use of thedifference of local weight between R and R, we choose the term with most difference as character term.(3) How to model user and filter information. According to the theory of fuzzy sets, a group of IF-THEN principals are constructed and implemented by ANFIS. The user's interests ui and Ui are stored in ANFIS as parameters. Pj is the input variable, and the relevance between Pj and ui, named Rpredj, is the output variable.(4) How to optimize and adjust the parameters. We adopt Candidate/Rank mode. The parameters are optimized in a way 'training-filtering-feedbacking-training-filtering'. The difference between the user's feedback Rusrj and the ANFIS output Rpredj is taken as the error. We optimize ri in a Widrow-Hoff algorithm, and optimize ui in a batch learning.According to the discussion before, the AUM&IF system, a prototype of user model-based Web filtering system, is evaluated by comparing its performance with analogous systems. The results achieved show that the use of user modeling techniques can improve the performance of Web information filtering system, and point out interesting challenges for future investigations.Some ideas in our work can be helpful to the similar application.