Research On Collaborative Filtering Recommendation Algorithm And Improvement

As one of the most successful personalized recommendation technologies until now, collaborative filtering technology obtains widely attention of researchers, and becomes the key issue need to be studied in the field of recommendation. This thesis has carried on the beneficial exploration and the research, mainly aiming at collaborative filtering technology, has brought in time forgetting technology, user preference and user feature, and proposed collaborative filtering algorithm which is based on the features of user attribute and the changes of user interest, precisely aiming at the defect of collaborative filtering technology. Moreover, this research has combined user interest, time effect, user preference for item and user feature into an organic unity, and made them supplement the merit mutually to improve the collaborative filtering algorithm which is based on user. The main work which the thesis has completed included:①To the recommendation system, this thesis has conducted the thorough research, including concept, research content and composition. In addition, it has introduced various recommendation technologies in details, and analyzed the superiorities and the insufficiencies of various types of recommendation technologies, as well as the characteristics of the existing typical recommendation system example. On the basis of this, the thesis lays emphasis on the research of the mainstream technology in the recommendation field---collaborative filtering technology, and makes a detailed and comprehensive introduction of collaborative filtering algorithm.②Considering the dynamic change of the user interest and drawing lessons from the forgetting rules of human being, this thesis has brought in forgetting technology, and put forward the improved nonlinear forgetting function on the basis of the current research achievement. According to the different temporal sequences it appears, this research entrusts each characteristics of user interest with different weight numbers, attenuates the item grading in different speeds according to time t, and change the degree of contribution of grading to recommendation results in different time.③Considering the effects of item properties to similarity, this research proposes the concepts of item properties'similarities, and takes the value factor of item itself into consideration. The similarities between item is not only determined by user's grading to merchandise item, instead, both item attribute and user grading can be synthesized to be used for measuring similarities between item. ④Considering that different item have different values for users, and the internal relations between item and users, this research put forward the concept of the degree of user preference. When the k-nearest neighbors are computed, the item can no longer be treated equally. Instead, the item should be entrusted with different weight numbers according to different preferences of users to item.⑤This thesis analyzes the reference value of user characteristics to recommendation process. In recommendation process, leads into the information of user characteristics reasonably, it is the expansion of k-nearest neighbors'model. The comparative experiment demonstrates that the recommendation quality is improved after bringing in the information of user characteristics, and the new user problem of electronic commerce website is solved to a certain extent.⑥Using the data base provided by MovieLens website, this thesis has conducted the tests on collaborative filtering algorithm based on user characteristics and interest changed. Besides, it has compared with other algorithms from many aspects, and verified the rationality and effectiveness of improving algorithm.