Research map / Probabilistic and scientific modeling / Causal discovery

Performative Validity of Recourse Explanations

NeurIPS 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Causal approaches to fairness · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that recourse explanations can invalidate themselves under their own performative effects, proves conditions for validity, and argues that only improvement-focused causal recourse (acting on causes) is robustly valid.

Problem

Recourse explanations are typically assumed to remain valid after model retraining, but when many applicants implement recommended actions, the induced distribution shift can change the decision boundary and render the recourse invalid (performativity of recourse).

Approach

Formalizes recourse-induced performative shifts in an SCM with an action variable, defines performative validity, and proves that recourse is invalid when actions are influenced by effect variables or intervene on direct effects of the target; advocates recourse acting only on causal variables (ICR).

Key results

• In the nonlinear additive setting, CE and CR decrease the conditional probability of a favorable outcome by 70% to 100%.
• In the GPA setting, the post-recourse acceptance rate drops by nearly 80% for CE and CR.
• ICR keeps the conditional distribution and acceptance rates essentially unchanged across settings, except ind. ICR in LCubic where the probability increases by 0% to 60%.

Measured effects reported in the paper

Surprising findings

• Recourse that only intervenes on causes (ICR) can still shift the conditional distribution when its actions are influenced by effect variables.
• Even under the common causal-cause-only assumption, a worked example shows recourse can be invalid when actions depend on effect variables.

Limitations

• Assumes causal sufficiency and no unobserved confounders, and access to the true SCM or causal graph.
• Results for ICR are proven only under specific assumptions (noise resampling or invertible aggregated noise); empirical ICR validity extends beyond the proven settings without a full characterization.
• Analysis is limited to single-step recourse with a one-time model refit.

Experimental setup

Datasets: GPA college admission dataset credit scoring dataset

Benchmarks: synthetic SCM settings (LAdd, LMult, NLAdd, NLMult, LCubic) GPA Credit

Metrics: pointwise difference in conditional probability of acceptanc post-recourse acceptance rate difference

Baselines: counterfactual explanations causal recourse (individualized) causal recourse (subpopulation-based)

Builds on: counterfactual explanations causal recourse performative prediction improvement-focused causal recourse

Closest papers by meaning

• Algorithmic recourse under imperfect causal knowledge: a probabilistic approach (NeurIPS 2020 · similarity 0.87)
• Improvement-Focused Causal Recourse (ICR) (AAAI 2023 · similarity 0.86)
• Algorithmic Recourse for Long-Term Improvement (ICML 2025 · similarity 0.83)
• Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance (AAAI 2024 · similarity 0.82)
• Predictive Performance Comparison of Decision Policies Under Confounding (ICML 2024 · similarity 0.82)
• A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.82)