Go to AAAI 2025 Workshop AICT homepageAdaptively Estimator Switching for Debiased Recommendation
Keywords: Recommendation, Selection Bias, Estimator Switching
Abstract: In the information era, recommendation systems play a crucial role in mitigating information overload by predicting user preferences based on historical interactions. However, traditional recommendation methods often neglect the issue of selection bias arising from non-random missing data, which compromises recommendation quality. To address this, existing approaches such as error imputation-based (EIB), inverse propensity scoring (IPS), and doubly robust (DR) estimators have been proposed. While these methods have demonstrated effectiveness, they suffer from limitations such as sensitivity to small propensity scores, high variance, and inaccuracies in error estimation. This paper introduces a novel switch estimator designed to flexibly integrate the strengths of EIB, IPS, and DR approaches while mitigating their respective weaknesses. Specifically, the proposed method employs a principled Monte Carlo sampling strategy to estimate relative errors in propensity scores and imputation, enabling adaptive threshold-based switching between estimators. This approach ensures robustness to issues arising from small propensity scores and large imputation errors. Experimental evaluations on three real-world datasets demonstrate the superior performance and robustness of the switch estimator in recommendation tasks. The proposed methodology advances the state-of-the-art by offering a practical and effective solution to selection bias in recommendation systems. We conduct experiments on three real-world datasets to show the effectiveness of our method.
Submission Number: 35
Loading