Incorporating latent space correlation coefficients to collaborative filtering

Collaborative Filtering (CF) is a popular approach to generate predicted rating of a target user on an item by aggregating neighbor users' ratings; these ratings are weighted by a correlation coefficient between two users. Thus, the user-user similarity computation is a significant step in CF to select proper neighborhood and exploit suitable correlation coefficients for prediction, and multiple weighting techniques have been proposed to enhance the performance. However, existing approaches compute the similarity directly based on users' rating vectors, which may lead the system to suffer from severe low-sparsity problem, and will also cause the system to be less interpretive because the rating only represents user's preference on a certain item but does not include extra feature information like attributes or genres. In this paper, we propose a method to compute the user' correlations in latent space by incorporating matrix factorization (MF) technique, and exploit the correlation coefficients in the prediction step of CF. We have evaluated the proposed approach with variant methods on MovieLens dataset to validate the effectiveness in CF.