Combining User-Based and Item-Based Models for Collaborative Filtering Using Stacked Regression

Collaborative filtering can be classified into user-based or item-based methods according to different assumptions. Our experiment results show that both user-based and item-based methods can reach a certain degree of accuracy, but the recommendation coverage of these two methods is significantly different. This paper qualitatively analyzes the advantages and disadvantages of such two methods, and found that user-based methods advantage in recommending popular items while item-based methods perform better at suggesting long tail items. Based on this analysis, we proposes a novel machine learning framework to systematically combine both. We train a model for each user-item pair for discovering the local preference of the user or item over both methods, and effectively combine user-based and item-based predictions using the preference information. Experiment results show that our approach can significantly improve the recommendation quality, and to a certain extend, alleviate the data sparsity problem.