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

Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Conformal prediction sets

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.

Proposes Stochastic Grouping Conformal Prediction (SGCP), which learns a latent stochastic grouping map and component score laws to build sample-specific local score transforms, reducing subgroup coverage gaps while keeping marginal validity and small prediction sets.

Problem

Standard conformal prediction gives only population-level coverage, so subgroups (e.g., in clinical data) can be under-covered; existing group-wise and localized methods either need sensitive attributes or become conservative and unstable with limited calibration data.

Approach

Learn a local score law for each sample as a posterior-averaged mixture of latent monotone calibration components (via a Gaussian latent state mapped to prototype-based memberships), then calibrate on the transformed (percentile) scores so samples with similar calibration behavior share evidence without predefined groups.

Key results

• SGCP achieves CovGap 4.2 on SYN, 11.8 on Nursery and 6.2 on BACH, the smallest among compared methods on those datasets (per text).
• On MIMIC-IV SGCP reports CovGap 0.8 versus FaReG's 0.5, with coverage 90.13% and AvgSize 1.107.
• SGCP reports the smallest or nearly smallest AvgSize across all four datasets, e.g., 2.224 on BACH versus 2.311 for RLCP.

Surprising findings

• Localized methods like RLCP and FaReG are competitive only in some settings with less uniform gains across benchmarks.
• Explicit subgroup-protection baselines reduce disparity at the cost of inflated coverage and larger prediction sets.

Limitations

• Gains on MIMIC-IV CovGap are behind FaReG (0.8 vs 0.5).
• Subgroup coverage guarantee depends on local score-law estimation error; no finite-sample bound on learned transform error is given beyond the oracle-assumption remark.

Experimental setup

Datasets: SYN (synthetic cardiopulmonary diagnosis) Nursery MIMIC-IV BACH

Benchmarks: SYN Nursery MIMIC-IV BACH

Metrics: overall coverage AvgSize CovGap

Baselines: Marginal CP Partial Equalized Exhaustive Equalized RLCP CluCP AFCP FaReG

Builds on: split conformal prediction RLCP FaReG AFCP

Closest papers by meaning

• Bridging Fairness and Efficiency in Conformal Inference: A Surrogate-Assisted Group-Clustered Approach (ICML 2025 · similarity 0.89)
• On the Burden of Achieving Fairness in Conformal Prediction (NeurIPS 2026 · similarity 0.89)
• Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery (NeurIPS 2026 · similarity 0.88)
• SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions (NeurIPS 2026 · similarity 0.87)
• Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage (ICML 2025 · similarity 0.87)
• Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands (NeurIPS 2025 · similarity 0.87)