counterfactual fairness analysis

Counterfactual prediction under evolving intervention policies or hypothetical decision scenarios remains challenging due to unobservable potential outcomes, hindering model identifiability, evaluation, and generalization. Method: We propose the first systematic theoretical framework addressing this challenge—comprising (i) identifiability conditions for counterfactual prediction models, (ii) a performance evaluation system targeting loss, AUC, and calibration, and (iii) robust hyperparameter selection under model misspecification. Our approach integrates causal inference principles, doubly robust estimation, and loss-driven evaluation metric design. Contribution/Results: Validated via simulation studies and a real-world clinical application—cardiovascular risk prediction in statin-naïve populations—the framework significantly improves out-of-distribution generalization and clinical decision reliability in counterfactual settings.

This paper addresses the problem of generating *implementable counterfactual explanations*: given a negative prediction from a machine learning model, the goal is to produce a multi-step sequence of feature interventions that leads to a positive outcome while strictly respecting inter-feature causal constraints. Methodologically, the authors propose CoGS—a novel framework that integrates causal graph modeling with Answer Set Programming (ASP). It encodes causal dependencies as logical rules and leverages ASP solvers to guarantee that each intervention step adheres to the underlying causal structure; a dedicated path-search algorithm further ensures computational efficiency and solution feasibility. Empirically, CoGS achieves superior performance across multiple benchmark datasets: its generated counterfactual paths are simultaneously causally valid, semantically plausible, and operationally feasible—outperforming existing single-step and causally agnostic approaches.

Current counterfactual prompting methods struggle to disentangle a large language model’s genuine sensitivity to a target variable from spurious sensitivity arising from superficial surface-form changes, leading to attribution bias. This work proposes a statistical comparison framework that introduces semantics-preserving paraphrastic perturbations as a baseline control. By contrasting the model’s output variations under targeted interventions against those induced by semantically equivalent rewrites, the approach enables a more accurate assessment of true model sensitivity. Integrating hypothesis testing, counterfactual editing, and multiple effect-size metrics, the method reveals that most reported effects on MedQA and MedPerturb datasets become statistically insignificant after baseline adjustment. However, it successfully detects significant gender bias in a professional biography generation task, demonstrating both its effectiveness and necessity.

This work formally defines and systematically investigates data poisoning attacks against counterfactual explanations (CEs), revealing their mechanisms for significantly inflating algorithmic redress costs at instance-, subgroup-, and population-levels. We propose theoretically grounded, provably correct poisoning strategies and conduct adversarial injection experiments on state-of-the-art CE generators—including DiCE and CFProto—complemented by a multi-level cost measurement framework. Results demonstrate that current SOTA CE methods exhibit systemic vulnerabilities: poisoning induces substantial increases in counterfactual path length or complete failure of redress generation. Our contribution is twofold: (1) it provides a critical security alert regarding the robustness of explainable AI systems; and (2) it establishes the first principled analytical paradigm for data poisoning targeting counterfactual explanations—thereby advancing rigorous robustness evaluation and defense research for trustworthy AI.

This study addresses the limitations of traditional control-based causal inference methods—such as matching and difference-in-differences—in settings characterized by pervasive or structurally ambiguous spillover effects, where reliance on uncontaminated control units impedes accurate identification of both average direct and spillover effects. Within the potential outcomes framework, this work provides the first systematic comparison between control-based and prediction-based counterfactual approaches—including interrupted time series and machine learning control—in terms of their identification capabilities. Through simulation and empirical analyses, the authors demonstrate that in environments with widespread interference, prediction-based methods can more reliably estimate certain causal parameters over short horizons, circumventing the stringent assumption of unperturbed units and thereby offering a promising alternative for causal inference under complex interference.

In high-stakes domains such as healthcare and criminal justice, it is often infeasible to re-conduct randomized controlled trials (RCTs) after updating machine learning models, rendering the causal effects of these updates on downstream outcomes—such as patient survival or recidivism rates—difficult to assess. This work proposes a novel partial identification approach that leverages historical RCT data and fine-grained relationships between prediction accuracy and downstream outcomes. Under two monotonicity assumptions—individual-level “counterfactual correctness” (i.e., correct predictions never lead to worse outcomes) and a trust relationship between subgroup predictive performance and outcomes—the method constructs tight bounds on the causal effect of the updated model. Simulations demonstrate that this approach yields more informative causal effect estimates compared to existing techniques.

Existing benchmarks for causal inference in time series are often restricted to observational data, limited in scale, or domain-specific, thereby hindering robust intervention and counterfactual analysis. This work proposes an open, extensible, and theoretically grounded framework for generating multivariate time-series structural causal models (TSCMs), which introduces several novel components: continuous-time intervention windows, counterfactual sampling with guaranteed positivity, mechanism-switching SCMs, and intervention profiles that embed trends and structural breaks. The framework also accommodates priors from causal foundation models. A training-scale dataset released under this framework comprises 100,000 trajectories spanning eight identifiable causal structures. Empirical results demonstrate that models trained on interventional data consistently and significantly outperform observation-only counterparts of equivalent capacity across all tested settings.