The Research Of Collaborative Filtering Recommendation Algorithm Based On Web Usage Mining

With the fast development of Internet and extensive applications of E-Commerce, more and more various information have flooded the net. Nowadays, how to get the right information that we need from the information sea has become one of the key issues for the researchers, experts as well as the Internet users. Under the circumstance of it,personalized recommendation systems comes into being and have attracted a great deal of research.There are many ways to actualize personalized recommendation. Among these ways, the collaborative filtering is a relatively more successful technology that is implemented in the systems today. However, as the system scale (such as the structure of the web site, the types of the items and the number of the customers) gradually becomes large, collaborative filtering algorithm has found its fatal limits, such as real-time requirement, data sparsity, and scalability problem.The traditional collaborative filtering recommendation systems demand the users to rate the items they use so as to get feedback. It frequently interrupts the users'browsing on the webs, even adversely affects the user's emotion. Furthermore, the fact that a considerable number of users are not willing to give rating comment undoubtedly makes the rating data sets extremely sparse. Aimed to settle the problems above, this paper will propose an approach to build a user-item matrix model based on web usage mining. By effectively mining web logs to discover some hidden and useful information(such as users, items, visiting times and visiting frequency), this method is able to acquire the rating marks of the users thus completing the data collection without breaking the normal browsing of the users. As the result of coverage rate of users'browsing webs is far more real and object than that of user rating, this method of data collection is better than the traditional one.To solve real-time and data-sparse problems of the traditional collaborative filtering, this paper will propose an improved collaborative filtering recommendation algorithm. Gathering the bilateral information of users and items in the original data set, we use the nearest neighbors of the users and items to predict the rating of unrated items so as to reduce the sparsity of data set. In addition, the clustering technology is introduced to cluster users of same type in the system and gather similar users together as much as possible. This step can be carried out off-line. As to the recommendation online, we only need to study the similarity between the target users and each user cluster center, then search nearest neighbors of the target users in the top certain number of the most similar user clusters. We can finally predict the rating of items according to the nearest neighbor set, thus generating recommendation to the target user. The proposed method not only reduces the researching scope of neighbors, but also improves the real-time of recommendation algorithm.Experiment has been made to prove that the proposed collaborative filtering algorithm is reasonable and effective. Compared to the traditional collaborative filtering algorithm, the proposed one can effectively solve the problem of data-sparse. Moreover, the experiment result shows that our algorithm is quite helpful since it offers higher rate of accuracy and higher speed of real-time, which is not shared by the traditional one.