Actively Semi-Supervised Collaborative Filtering

Abstract

Collaborative filtering (CF) has been widely used in various recommender systems, but often suffers from the problem of data sparsity which dramatically degrades the recommendation performance. In this paper, we propose a co-training style semi-supervised CF approach towards the task of rating prediction, which exploits a few observed ratings in conjunction with copious unobserved ones to reduce sparsity. In each round of co-training iterations, our approach utilizes two different neighborhood-based recommenders, each of which labels the unobserved data for the other recommender; in particular, the most informative unobserved examples are actively selected for labeling, and then the labeling confidence is estimated through validating the influence of the labeling of unobserved examples on the observed ones. Experiments results on the three datasets demonstrate that our approach can effectively exploit unobserved data to improve CF predictions, and achieves better performance than other counterparts.