Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback

摘要: Recommender systems widely use implicit feedback such as click data because of its general availability. Although the presence clicks signals users' preference to some extent, lack does not necessarily indicate a negative response from users, it is possible that users were exposed items (positive-unlabeled problem). This leads difficulty in predicting preferences feedback. Previous studies addressed positive-unlabeled problem by uniformly upweighting loss for positive or estimating confidence each having relevance information via EM-algorithm. However, these methods failed address missing-not-at-random which popular frequently recommended are more likely be clicked than other even if user have considerable interest them. To overcome limitations, we first define an ideal function optimized realize recommendations maximize and propose unbiased estimator loss. Subsequently, analyze variance proposed further clipped includes special case. We demonstrate expected improve performance recommender system, considering bias-variance trade-off. conduct semi-synthetic real-world experiments method largely outperforms baselines. In particular, works better rare less observed training data. The findings can achieve objective recommending with highest relevance.