A Second-Order Markov Random Walk Approach for Collaborative Filtering

Collaborative filtering is the most widely used technique to generate recommendations for an active user by the opinions of the others. However, the challenge is that sometimes the data set is too sparse to identify the similarities of user interests. Random walk on bipartite graphs has been proposed to solve this problem. By exploring transitive association through the first-order Markov process, it is able to find a group of like-minded users for an active user, even if they have no co-rated items. It works for the ratings in binary, but quite often people rate items with numerical scale (e.g. 1-5), which makes it hard to be applied. In this paper, we propose a second-order Markov process to overcome the limitation. Experimental results demonstrate that this approach outperforms the classic collaborative filtering methods with substantial improvements in prediction accuracy and coverage on sparse data set.