Multi-order nearest neighbor prediction for recommendation systems

Abstract

Recommendation systems (RSs) can establish a relationship between users and items and recommend the seemingly unrelated but actually interesting items to target users by utilizing their behavior. However, the recommendation performance of RSs is seriously affected by the issue of data sparsity because the information amount that each user involves is very limited with the continuous growth of both users and items. To solve the issue of data sparsity in RSs, a novel collaborative filtering prediction method is proposed, called multi-order nearest neighbor prediction (MNNP). The concept of multi-order nearest neighbors is an extension of friends of a friend. By using multi-order nearest neighbors, MNNP can not only expand the range of neighbors but also implement the neighbor propagation efficiently. For a target user, MNNP needs to search its multi-order nearest neighbors and successively uses them to iteratively update the rating matrix. To show the procedure of MNNP, we illustrate it by an example. Extensive experiments on real-world datasets show that MNNP has a good performance.