Adaptive Calibrated Doubly Robust Estimators for Debiased Recommendation

Recommender systems (RS) aim to provide personalized content, but they face a significant challenge in unbiased learning due to selection bias, as users typically interact only with items they might prefer. To address this issue, various methods such as Inverse Propensity Scoring (IPS) and Doubly Robust (DR) estimators have been proposed. However, these methods cannot guarantee the accurate calibration of their propensity models. Therefore, some enhanced DR methods, such as DCE-DR and Cali-MR, are proposed to address this issue, but they commonly adopt a pre-defined binning strategy to empirically estimate the calibration error. We formalize this partition dependence as “bin-sensitivity” and show that calibration estimates can vary substantially across equally reasonable partitions, undermining reliability. To address this problem, we propose a robust calibration framework that, unlike existing fixed-partition methods, adopts a minimax optimization strategy to minimize the worst-case calibration error across all possible partitions. In practice, we develop an efficient greedy algorithm to approximate this worst-case partition while maintaining computational tractability. Extensive experiments on three real-world datasets demonstrate the effectiveness of our proposed method. The source code is publicly available at https://github.com/Qshuqi/AdapCDR