Bilateral Self-unbiased Recommender Learning for Missing-not-at-Random Implicit Feedback

Unbiased recommender learning aims at eliminating the intrinsic bias from implicit feedback under the missing-not-at-random (MNAR) assumption. Existing studies primarily focus on estimating the propensity score for item popularity bias but neglect to address the exposure bias of items caused by recommender models, i.e., when the recommender model renders an item more frequently, users tend to more click the item. To resolve this issue, we propose a novel unbiased recommender learning framework, namely Bilateral self-unbiased recommender (BISER). Concretely, BISER consists of two parts: (i) estimating self-inverse propensity weighting (SIPW) for the exposure bias during model training and (ii) utilizing bilateral unbiased learning (BU) to minimize the difference for model predictions between user- and item-based models, thereby alleviating the high variance from SIPW. Our extensive experiments show that BISER significantly outperforms state-of-the-art unbiased recommender models on various real-world datasets, such as Coat, Yahoo! R3, MovieLens-100K, and CiteULike.