An Improved Collaborative Filtering Algorithm Based On User And Item Clustering

With the rapid development of Internet, human beings have stepped into information-overloaded age from the information-lacking age. In order to overcome the challenge brought by information-overloaded age, lots of new technologies have been invented, such as search engine and recommendation system. Now the technology of search engine has been so powerful that we can find out almost everything we want, but it has to wait for our typing. The emergence of recommendation system has made up the limitation of search engine, by analyzing the history data of users. It can figure out the interests of users, and then initiatively recommend the things that users might like. Nowadays, recommendation system has been developing into an extremely important part of e-commerce websites, because it can offer not only good experience to users, but also bigger profits to the service providers.Among large quantities of recommendation system, collaborative filtering is the most popular. Lots of large e-commercial websites are using collaborative-filtering-based recommendation system, such as Amazon, Google and so on.Traditional collaborative recommendation algorithms predict user’s preference by the ratings that other users rated items. But it ignores the information of users’and items’ properties, so the clustering result is not quite accurate. To generate a user’s closest neighbors, it has to compute the similarities with all the other users, so it consumes a lot of time. And it is quite sensitive to the sparsity of user-item rating matrix, especially when the sparsity is quite low, and it is not quite accurate.By using the technology of clustering, the similar users can be divided into the same cluster. So we just need to compute the similarity value of two users in the same cluster. In this way, the performance of recommendation system can be improved.In order to raise the accuracy in the process of clustering users and items, a new algorithm is proposed in this paper. In this algorithm, both the properties of items and the ratings of items are used to compute the similarity value. When the result of item clustering is generated, user-item clustering matrix is used to generate the clustering of users. In the end, by comparing the experiment data, the effectiveness of the new algorithm can be verified.