Local Top-N Recommendation via Refined Item-User Bi-Clustering

Abstract

Top-$N$ recommendation has drawn much attention from many portal websites nowadays. The classic item-based methods based on sparse linear models (SLIM) have demonstrated very good performance, which estimate a single model for all users. Lately, local models have been considered necessary since a user only resembles a group of others but not all. Moreover, we find that two users with similar tastes on one item group may have totally different tastes on another. Thus, it is intuitive to make preference predictions for a user via item-user subgroups rather than the entire feedback matrix. For elegant local top-$N$ recommendation, this paper introduces a bi-clustering scheme to be integrated with SLIM, such that item-user subgroups are softly constructed to capture subtle preferences of users. A novel localized recommendation model is hence presented, and an alternative direction algorithm is devised to collectively learn item coefficient for each local model. To deal with the data sparsity issue during clustering, we conceive a refined feature-based distance measure to better model and reflect user-item interaction. The proposed method is experimentally compared with state-of-the-art methods, and the results demonstrate the superiority of our model.