Research map / Learning theory and algorithms / Prediction and decision losses

Performative Prediction with Selective Labels

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Data attribution and recourse

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.

Formalizes performative prediction with selective labels (threshold-based observation), shows naive accept-only retraining can diverge, and proposes Robust RRM using worst-case strongly convex upper-bound loss with label-probability intervals tightened from accepted data.

Problem

Repeated risk minimization in performative prediction assumes full label access after deployment, but under selective labels (labels observed only for accepted applicants), retraining on observed data can prevent convergence to the performatively stable solution.

Approach

Replace the unobservable loss on rejected samples with a strongly convex worst-case upper bound (log-sum-exp over label-probability interval endpoints) that does not depend on the true label, and tighten the intervals over time using sensitivity of the label distribution and previously accepted samples.

Key results

• Robust RRM with learned label-probability limits reduces distance to the stable solution from about 1.5 (trivial limits) to 0.212 and closely matches Oracle RRM performative risk (0.336).
• Performative gap falls below 1e-4 across sensitivity levels epsilon in {50, 90, 150}.
• DR Estimator diverged with very high performative risk and disparity under thresholded decisions, while Accept-only had the largest distances among stable methods.

Surprising findings

• Underestimating L_x produced a smaller empirical distance to the stable solution than correct specification.
• Accept-only retraining can end arbitrarily far from the stable solution even when no performative effects exist.

Limitations

• Guarantees are population-level; no finite-sample bounds.
• Sensitivity constants L_x and L_theta are not identifiable from accepted samples under deterministic thresholds and require domain knowledge or exploration.
• Focuses on deterministic threshold policies; stochastic policies not covered.

Experimental setup

Datasets: FICO credit score dataset (initial distribution) simulated lending environment

Benchmarks: simulated fair lending environment

Metrics: performative risk demographic parity distance to stable solution theta^PS performative gap

Baselines: RRM with Oracle label access Accept-only RRM Affine RM DCEM DR Estimator

Builds on: performative prediction (Perdomo et al. 2020) partial identification decision-dependent distribution sensitivity Affine risk minimization DCEM doubly robust estimator

Closest papers by meaning

• Tight Lower Bounds and Improved Convergence in Performative Prediction (NeurIPS 2025 · similarity 0.87)
• Performative Prediction with Bandit Feedback: Learning through Reparameterization (ICML 2024 · similarity 0.86)
• On the Impact of Performative Risk Minimization for Binary Random Variables (ICML 2025 · similarity 0.86)
• Active, anytime-valid risk controlling prediction sets (NeurIPS 2024 · similarity 0.86)
• Uncertainty Herding: One Active Learning Method for All Label Budgets (ICLR 2025 · similarity 0.86)
• Optimal Classification under Performative Distribution Shift (NeurIPS 2024 · similarity 0.85)