Moment Similarity of Random Variables to Solve Cold-start Problems in Collaborative Filtering
Hyeong-Joon Kwon, Kwang‐Seok Hong · 2009
The cold-start problem is a primary factor causing performance loss in collaborative filtering. In this paper, we examine a fatal flaw of existing similarity measures in the cold-start condition. We propose a novel method, MSRV, using the moment of a random variable to solve the weaknesses of existing similarity measures that contain vector cosine similarity and correlation analysis-based methods. The proposed method is based on a prudent concept; if the expectation of the difference between two random variables is low, they will be similar to each other. We improve memory-based collaborative filtering performance using the moment that is a major statistical parameter. An experiment using various datasets confirms that the proposed method demonstrates significantly improved prediction performance compared to existing measures in full rating experiments.