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

Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Conformal prediction sets · 0 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.

Proposes Self-Organized Conformal Prediction (SOCP), which trains a Self-Organizing Map on unlabeled inputs and calibrates conformal thresholds from BMU-cell or neighborhood buffers to reduce regional coverage gaps while keeping predictor and score fixed.

Problem

Marginal conformal coverage from pooled calibration can hide systematic undercoverage in heterogeneous regions of the input space, including safety-critical subgroups.

Approach

Use a SOM trained without calibration labels to define input-space groups; at test time, compute the conformal quantile from the query's BMU cell, a fixed grid neighborhood, or a prototype-distance enlargement (with a training-routed global budget), with finite-sample validity for retrieved sets and an explicit KS bias for central-cell interpretation.

Key results

• Across ten datasets, SO-SCP reduces the weighted coverage gap relative to pooled SCP on nine datasets, with a mean relative change of -14.3% at a mean output-size change of +3.0%.
• Under fixed-neighborhood retrieval, SO composition lowers the ten-seed mean WCovGap in 43 of 50 dataset-score comparisons.
• SO-SCP lowers WCovGap on average over paired seeds for every dataset at external partition granularities K=10, 25, and 50 (30 of 30 comparisons).

Surprising findings

• Mondrian k-means lowers WCovGap more than SO-SCP on all five classification datasets, though it enlarges sets more on four of them.
• Lcp and RLCP return unbounded intervals on sparse regression datasets (up to 28.5% of predictions), while SO-SCP keeps all intervals finite on Bike Sharing, Bio, and California Housing.

Limitations

• Locality depends on the representation, map resolution, and retrieval rule.
• Sparse calibration buffers yield infinite cutoffs under fixed neighborhoods; Regime 3 is needed for small calibration sets.
• Central-cell guarantees pay an explicit KS bias term that is not guaranteed to be small.

Experimental setup

Datasets: CASP (Bio) Bike Sharing California Housing Concrete Compressive Strength Auto MPG CIFAR-10 CIFAR-100 MNIST Covertype ImageNet-1k

Benchmarks: CASP (Bio) Bike Sharing California Housing Concrete Compressive Strength Auto MPG CIFAR-10 CIFAR-100 MNIST Covertype ImageNet-1k

Metrics: WCovGap marginal coverage output size (interval width / set size) worst-slab conditional coverage (WSC) SSC gap L1-ERT

Baselines: split conformal prediction (pooled SCP) LCP RLCP Mondrian k-means Classwise calibration Clustered calibration CQR CHR NFCP RCP APS RAPS

Builds on: split conformal prediction Mondrian conformal prediction localized conformal prediction conformalized quantile regression least ambiguous set-valued classifiers (LAC)

Closest papers by meaning

• Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability (NeurIPS 2026 · similarity 0.88)
• Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage (ICML 2025 · similarity 0.88)
• On the Burden of Achieving Fairness in Conformal Prediction (NeurIPS 2026 · similarity 0.86)
• Bridging Fairness and Efficiency in Conformal Inference: A Surrogate-Assisted Group-Clustered Approach (ICML 2025 · similarity 0.85)
• CONTRA: Conformal Prediction Region via Normalizing Flow Transformation (ICLR 2025 · similarity 0.85)
• SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions (NeurIPS 2026 · similarity 0.85)