adversarial ips weighting

In causal effect estimation, historical decision policies induce distributional shifts between treatment and control groups, causing inverse probability weighting (IPW) to suffer from instability due to propensity score estimation bias and extreme weights. This paper proposes a distributionally robust causal estimation framework that—uniquely—integrates distributionally robust optimization (DRO) with weighted Rademacher complexity regularization: DRO mitigates ambiguity in propensity score estimation, while the regularization curbs statistical variance of weighted estimators. We approximate the DRO objective via a computationally tractable adversarial loss, preserving theoretical rigor while enhancing practicality. Extensive experiments on multiple synthetic and real-world datasets demonstrate that our method consistently outperforms state-of-the-art IPW and DRO baselines across estimation accuracy, robustness to distributional shift, and estimator stability.

To address exposure bias in implicit feedback for recommender systems, this paper proposes a counterfactual learning framework that integrates inverse propensity scoring (IPS) weighting with an enhanced Bayesian personalized ranking (BPR) objective. The method mitigates high variance induced by extreme IPS weights via a novel propensity regularization term and improves offline evaluation stability and robustness through a self-normalized IPS estimator. It unifies IPS-weighted training, regularized BPR optimization, direct estimation, and multiple evaluation variants within a single coherent framework. Experiments on synthetic data and MovieLens 100K demonstrate that the proposed approach significantly reduces evaluation variance—by an average of 37%—and enhances model generalization under unbiased exposure settings. The framework thus achieves a favorable trade-off between training effectiveness and evaluation reliability.

本文提出一种方法，通过结合因果推断的敏感性分析与Wasserstein分布鲁棒优化，解决因未观测混杂因素导致的风险评分学习问题。

This paper investigates the impact of covariate balancing on average treatment effect (ATE) and average treatment effect on the treated (ATT) estimation, and establishes the theoretical foundation for numerical equivalence among inverse probability weighting (IPW), augmented IPW (AIPW), and IPW-regression adjustment (IPWRA) estimators. We rigorously prove that when propensity scores are estimated via covariate balancing methods—such as inverse probability tilting (IPT) for ATE or covariate-balancing propensity score (CBPS) for ATT—the three estimators are algebraically identical, and their weights are automatically normalized. This equivalence unifies the theoretical frameworks of weighted and doubly robust estimation, substantially improving finite-sample stability and precision. Moreover, the result enables a novel analytical pathway for identifying the local average treatment effect (LATE) under unmeasured confounding, thereby advancing model robustness and computational consistency in causal inference.

This paper addresses the robustness challenge of average treatment effect on the treated (ATT) estimation under the no-unconfoundedness assumption when high-dimensional covariates or poor overlap undermine conventional methods. We propose a finite-information aggregation estimation framework situated between Manski’s bounds and inverse probability weighting (IPW). Our approach integrates a constrained dependence function design with a variant of IPW to jointly achieve robustness against model misspecification and efficiency in information utilization. We establish asymptotically valid interval estimation theory for the resulting estimator. Simulation studies and empirical applications demonstrate that the proposed method substantially tightens the identification bounds, exhibits superior robustness to increasing covariate dimensionality and overlap deficiency, and consistently outperforms both classical Manski bounds and standard IPW estimators in finite-sample performance.

本文提出了一种自适应双重鲁棒方法(ADR)，结合自适应重要性加权和奖励回归，以解决排名策略的离线评估问题，特别是在用户行为多样的情况下。

This study addresses the bias in causal effect estimation arising from misspecification of the propensity score model in inverse probability weighting (IPW). To mitigate this issue, the authors propose two clustering-informed strategies: a cluster-augmented IPW approach and a global propensity score model incorporating cluster membership indicators. The robustness of these methods is systematically evaluated through Monte Carlo simulations and an empirical analysis of breast cancer data across varying sample sizes and model specifications. Results demonstrate that the proposed clustering-aware methods substantially reduce both estimation bias and mean squared error, particularly when latent subgroup structures are present. Furthermore, they enable subgroup-specific causal effect estimation and significantly enhance robustness against propensity score model misspecification.

本文提出两种基于倾向评分的算法，用于在观察研究中估计平均处理效应，同时保护数据隐私，减少了误差和偏差。

In observational studies with limited covariate overlap, conventional inverse probability weighting (IPW) estimators of the average treatment effect (ATE) often exhibit substantial bias and unreliable confidence intervals due to violations of the strong overlap assumption. This work proposes a robust IPW approach based on polynomial extrapolation: by constructing a sequence of surrogate estimators indexed by a tuning parameter, it fits a polynomial function to these estimates and extrapolates to the target ATE, thereby substantially reducing reliance on strong overlap while preserving the original ATE definition. Theoretical analysis establishes that the proposed estimator remains consistent and asymptotically normal under weaker overlap conditions. Simulation studies demonstrate that it achieves markedly improved estimation accuracy and confidence interval coverage compared to standard IPW.

This study addresses the interpretational ambiguity of weighted estimators when treatment effects are heterogeneous, as their validity hinges critically on the choice of weights. To tackle this issue, the authors propose an estimator that minimizes worst-case bias and construct confidence intervals that are uniformly valid over a broad class of weighting schemes. Their approach integrates minimax bias reduction, bounds from heterogeneity-robust sensitivity analysis, and theoretical characterizations of discrepancies among weighted estimators, thereby enabling inference robust to weight uncertainty. Empirical applications illustrate the method’s utility: in Lakdawala et al.’s event study, findings remain robust across a wide range of weights, whereas in the Project STAR experiment, conclusions prove sensitive even to minor perturbations of baseline weights.