Semi-supervised collaborative filtering ensemble

Abstract

Collaborative filtering (CF) plays a central role in recommender systems, but often suffers from the data sparsity issue that dramatically degrades the recommendation performance. In this paper, we propose a Semi-Supervised Ensemble Filtering (SSEF) method to improve the recommendation performance by assembling three popular CF techniques in a co-training framework. Concretely, SSEF first initializes three weak predictors with labeled examples by three different CF algorithms independently. Two predictors generated by neighborhood methods are then merged, along with the remaining one generated by latent factor model, serve as two base recommenders, each of which labels the unlabeled examples for the other recommender during the co-training process. To exploit unlabeled data safely, the labeling confidence is estimated by validating the influence of the pseudo-labeled examples on the labeled ones. The final prediction is made by blending the outputs from the three predictors enhanced with unlabeled data. Extensive experiments on three public benchmarks demonstrate the effectiveness of the proposed SSEF by comparing to a number of state-of-the-art CF techniques, including semi-supervised, ensemble, and side-information based solutions.