Uncertainty-Aware Debiased Recommendation: A Bayesian Doubly Robust Perspective

Recommender systems are inherently plagued by selection bias arising from the Missing Not At Random (MNAR) nature of observed interactions. Doubly Robust (DR) learning, which integrates Inverse Propensity Scoring (IPS) and Error Imputation-Based (EIB) models, has emerged as a cornerstone for bias mitigation. However, traditional DR methods rely heavily on the precise point estimates of propensities and imputations, making them highly susceptible to catastrophic variance explosion under model misspecification or extreme data sparsity. To address these limitations, we propose the Bayesian Doubly Robust (BDR) learning framework. This framework shifts the debiasing paradigm from rigid point estimation toward robust distributional inference by leveraging Monte Carlo Dropout to capture epistemic uncertainty and introducing an Entropy Tilting mechanism. By minimizing Kullback-Leibler (KL) divergence within the function space, BDR dynamically calibrates posterior samples to satisfy causal unbiasedness moment conditions, thereby rectifying imputation errors without the need for model retraining. Theoretical analysis demonstrates the asymptotic unbiasedness of BDR even when the "accurate imputation assumption" is relaxed. Extensive evaluations on the Coat, Yahoo!, and KuaiRec datasets confirm that BDR excels in extremely sparse scenarios and functions as a model-agnostic, "plug-and-play" framework that consistently enhances the performance of state-of-the-art (SOTA) models.