A unified latent factor correction scheme for collaborative filtering

Collaborative filtering is the most popular technique to ease the information overload issue in the field of recommender system. The nearest neighbor based method and the latent factor based model are two widely used collaborative filtering methods. In order to benefit from both approaches, some researchers have proposed strategies to combine them, and the combinations have been shown to obtain more accurate results, especially during the Netflix competition. However, the unified scheme, which uses the neighborhood information to correct the learnt latent factors, is not well researched. In this paper, we generalize a novel unified scheme by correcting the latent features of users and items with the neighborhood information to boost the recommendations. We further elaborate several state-of-the-art latent factor models and some relationship integrating strategies into the proposed scheme. Finally, we conduct several series of experiments to compare the performance of different methods and latent factor based models within the unified scheme, and conclude with some suggestions in deploying the recommender systems.