PU Active Learning for Recommender Systems

In recommender systems, supervised information is usually obtained from the historical data of users. For example, if a user watched a movie, then the user-movie pair will be marked as positive. On the other hand, the user-movie pairs did not appear in the historical data could be either positive or negative. This phenomenon motivates us to formalize the recommender task as a Positive Unlabeled learning problem. As the model trained on the biased historical data may not generalize well on future data, we propose an active learning approach to improve the model by querying further labels from the unlabeled data pool. With the target of querying as few instances as possible, an active selection strategy is proposed to minimize the expected loss and match the distribution between labeled and unlabeled data. Experiments are performed on both classification datasets and movie recommendation dataset. Results demonstrate that the proposed approach can significantly reduce the labeling cost while achieving superior performance regarding multiple criteria.