Transfer learning for one-class recommendation based on matrix factorization

Abstract

The One Class Recommender System aims at predicting users future behaviors according to their historical actions. In these problems, the training data usually only contains binary data which reflects behavior that has or has not happened. Thus, the data is sparser than traditional rating prediction problems. There are two current ways to tackle the problem: first, using knowledge transferred from other domains to mitigate the data sparsity problem and second, providing methods to distinguish negative data and unlabeled data. However, it is not easy to transfer knowledge simply from a source domain to target domain since their observations may be inconsistent. In addition, without data from an external source, distinguishing negative and unlabeled data is sometimes infeasible. In this paper, we propose a novel matrix tri-factorization method to transfer useful information from the source domain to the target domain. Then we embed this method into a cluster-based SVD (singular value decomposition) framework. In several real-world datasets, we show our method achieves better prediction precision than other state-of-the-art methods. To date, the cluster-based SVD method has been on an online shopping site for two months, and its performance (conversion rate in sales) is rating among the best.

Date of Award	2015
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology