Confidence-Learning Based Collaborative Filtering with Heterogeneous Implicit Feedbacks

Abstract

Implicit feedbacks, which indirectly reflect opinions through observing user behaviors, have recently received more and more attention in recommendation communities due to their accessibility and richness in real-world applications. Most of the existing implicit-feedback-based recommendation algorithms only exploit one type of implicit feedback. In real-world applications, there is usually more than one type of implicit feedback. Considering the sparsity problem of recommender systems, it is significant to leveraging more available data. In this paper, we study the heterogeneous implicit feedbacks problem, where more than one type of implicit feedback is available. We study the characteristics of different types of implicit feedbacks, and propose a unified approach to infer the confidence that we can believe a user prefers an item. Then we apply the inferred confidence to both point-wise and pair-wise matrix factorization models, and propose a more generic strategy to select training samples for pair-wise methods. Experiments on real-world e-commerce data show that our methods outperform the state-of-art approaches, considering several commonly used ranking oriented evaluation criterions.