Robust Personalized Recommendation under Hidden Confounding in MNAR

2026-05-20

AI Summary

This work addresses selection bias in recommender systems caused by unobserved confounders and missing-not-at-random (MNAR) data by proposing a Personalized Unobserved Confounding-aware Interaction Debiasing framework (PUID). PUID introduces, for the first time, user-item-level heterogeneous sensitivity bounds that relax the conventional assumption of global homogeneity. By integrating adversarial optimization with a pre-trained model-guided mechanism (BPUID), the framework achieves robust debiasing without requiring randomized controlled trial data. Experimental results demonstrate that PUID significantly outperforms existing global debiasing methods across three real-world datasets, effectively enhancing both accuracy and stability of recommendations under hidden confounding.

Abstract

Recommender systems often rely on observational user--item interaction data, which is prone to selection bias due to users' selective interactions with items. Inverse propensity weighting and doubly robust estimators effectively mitigate selection bias under observed confounding, but are unreliable in the presence of hidden confounders. Existing approaches relying on randomized controlled trials (RCTs) or global sensitivity bounds are constrained in practice: RCTs demand costly experimental data, while global sensitivity bounds presume a uniformly bounded effect of unmeasured confounders on propensities through sensitivity analysis, thereby neglecting heterogeneity across user--item interactions. To overcome this limitation, we propose a novel framework, which estimates user--item level sensitivity bounds, thereby substantially relaxing the homogeneity assumption inherent in global sensitivity bounds named Personalized Unobserved-Confounding-aware Interaction Deconfounder (PUID). To ensure both robustness and predictive accuracy, we further develop an adversarial optimization strategy and propose a benchmark-guided variant (BPUID) that incorporates pre-trained models as stabilizing references. Extensive experiments on three real-world datasets demonstrate that our approach significantly outperforms global methods under hidden confounding, without requiring RCT data.

Problem

• hidden confounding
• selection bias
• recommender systems
• sensitivity analysis
• MNAR

Innovation

• hidden confounding
• personalized sensitivity bounds
• adversarial optimization
• MNAR
• deconfounding