Improve Performance of Collaborative Filtering Systems Using Backward Feature Selection

In E-Commerce, recommender systems recommend to users those items that may they like and have not been seen by them. One of the most popular techniques in recommender systems is collaborative filtering. For the current user, collaborative filtering technique uses the rating information of similar users to recommends new items. Therefore, clustering methods can be performed to divide users into k clusters as users of each cluster are similar. In this paper, k-means clustering method is used to cluster users based on user-item matrix. But, the sparsity of user-item matrix leads to produce incorrect similarity values between different users. To decrease this undesirable effect, removing some of the items with much more zero rates than the other items, can be done before clustering method. By eliminating these features, k-means clustering produces clusters which include more similar users. In this paper, backward feature selection is used to remove the irrelevant features of user-item matrix. Simulation results on MovieLens dataset indicate that this method improves the performance of the recommender system.