Boosting collaborative filtering with an ensemble of co-trained recommenders

Abstract

Collaborative Filtering (CF) is one of the best performing and most widely used approaches for recommender systems. Although significant progress has been made in this area, current CF methods still suffer from cold-start and sparsity problems. A primary issue is that the fraction of users willing to rate items tends to be very small in most practical applications, which causes the number of users and/or items with few or no interactions in recommendation databases to be large. As a direct consequence of ratings sparsity, recommender algorithms may provide poor recommendations (reducing accuracy) or decline recommendations (reducing coverage). This paper proposes an ensemble scheme based on a co-training approach, named ECoRec, that drives two or more recommenders to agree with each others’ predictions to generate their own. The experiments on eight real-life public databases show that better accuracy can be obtained when recommender algorithms are simultaneously trained from multiple views and combined into an ensemble to make predictions.