User-based collaborative filtering algorithm fusing the local and global nearest neighbour

Abstract

Collaborative filtering method is one of the most successfully used recommendation algorithm. It can be divided into the global nearest neighbour based method and the local nearest neighbour based method. The global neatest neighbour based method has high stability but the recommended results are not sufficiently personalised. The local nearest neighbour based method is more accurate and personalised than the global neatest neighbour based method, but the prediction failure rate is high and it's not stable. Aiming at these problems, a new user-based collaborative filtering algorithm fusing the local and global nearest neighbour is proposed. Firstly, the max-min K-medoids clustering algorithm is applied for clustering items into several clusters. The local similarity of users is calculated in each item cluster. Secondly, the factor of correlation-weighted is introduced to improve the accuracy of the global similarity among users. Finally, this new similarity among users is presented to optimise the selection of neighbours of target user. This new method increases the recommendation precision while avoiding the shortcoming of instability. The experimental results on EachMovie and MovieLens-100K show that the prediction reliability and accuracy of this new method are better than traditional ones.