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

Hierarchical Conformal Classification

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 Hierarchical Conformal Classification (HCC), which builds prediction sets of mixed-level nodes in a class DAG by searching over non-overlapping leaf covers and calibrating a conformal predictor per cover, with Bonferroni correction at inference.

Problem

Standard conformal prediction ignores class hierarchies, producing flat prediction sets that are not compact or semantically specific; existing hierarchical CP methods either are limited to trees, lack leaf-coverage guarantees, or scale poorly.

Approach

Formulates prediction-set construction over a label DAG as a constrained optimization (minimize set size plus beta-weighted covered leaves subject to leaf coverage), shows that only non-overlapping leaf covers need to be searched, propagates scores and labels upward, calibrates one conformal predictor per cover, and prunes candidates dynamically at inference.

Key results

• On img (1-alpha=0.98), HCC reduces prediction set size from 11.04 (standard CP) to 4.41 while meeting coverage (0.9846) and lowering cost to 8.09 vs 11.49.
• In a user study of 500 samples, annotators preferred HCC over standard CP in about 57% (dbp) and 71% (img) of cases (binomial test significant).
• On dbp at 1-alpha=0.9997, HCC attains coverage 0.9999 with average prediction set size 1.25 versus 2.67 for standard CP.

Surprising findings

• HCC covers more leaves than standard CP on average (e.g., 92 vs 11 on img), a moderate cost for compactness.
• Users sometimes preferred standard CP, e.g., in dbp abstract hierarchy labels were less helpful; ImageNet multi-object images led to abstract ancestor labels like 'object'.

Limitations

• Bonferroni correction over many NOL-covers is conservative and can inflate prediction set sizes.
• Moderate user study with fair-to-moderate inter-annotator agreement (Fleiss kappa 0.33-0.53) and authors as annotators in Appendix F.
• The number of NOL-covers can grow large for big taxonomies, requiring limited-search heuristics.

Experimental setup

Datasets: DBpedia14 ImageNet1K GTZAN

Benchmarks: DBpedia14 (dbp) ImageNet1K (img) GTZAN (gtz)

Metrics: coverage cost prediction set size number of covered leaves user preference

Baselines: Standard CP LCA Structured Conformal Prediction (SCP)

Builds on: split conformal prediction learn-then-test conformal risk control

Closest papers by meaning

• Generative Conformal Prediction with Optimized Coverage Allocation (NeurIPS 2026 · similarity 0.90)
• Conformal Structured Prediction (ICLR 2025 · similarity 0.89)
• Optimal transport-based conformal prediction (ICML 2025 · similarity 0.88)
• Conformal Prediction with Learned Features (ICML 2024 · similarity 0.88)
• Valid Selection among Conformal Sets (NeurIPS 2025 · similarity 0.88)
• Direct Prediction Set Minimization via Bilevel Conformal Classifier Training (ICML 2025 · similarity 0.88)