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

FedCF: Fair Federated Conformal Prediction

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.

Extends Conformal Fairness (CF) to federated learning via FedCF, using a mixture-of-clients decomposition of group-conditional coverage gaps, a descent-based threshold search, and two aggregation protocols, evaluated on ACS, Pokec, and Fitzpatrick datasets.

Problem

Conformal prediction sets can give unequal conditional coverage across sensitive groups, and existing fair CP methods assume centralized data, while federated settings add heterogeneity, communication, and privacy constraints.

Approach

Decompose group-conditional coverage into per-client terms computable locally under partial exchangeability, aggregate them on the server to bound coverage gaps, and search for the minimal threshold satisfying a closeness criterion with a descent-based update instead of grid search.

Key results

• On Fitzpatrick with RAPS and 8 clients, FedCF at c=0.1 reduces demographic parity disparity from 0.166 (base FCP) to 0.089 while meeting the bound.
• On ACSEducation, using group-and-class-wise thresholds lowers average set size versus classwise thresholds (e.g., 4.391 vs 4.823 at c=0.15).
• Descent-based search reaches comparable fairness-efficiency results with about 100 communication rounds versus 1000 for grid search, roughly a 10x reduction.

Surprising findings

• MLE point estimates can exceed the closeness criterion yet are fairer than baseline while sacrificing less efficiency than interval bounds.
• Enforcing fairness on the ACS and Fitzpatrick settings can produce large set-size increases, with a degenerate near-full-efficiency predictor being trivially fair.

Limitations

• Relies on partial exchangeability, which may be violated under distribution drift.
• Interval bounds give conservative, inefficient (large) prediction sets; MLE point estimates trade guarantees for efficiency.
• Requires sufficient calibration data per group-label pair across the federation.

Experimental setup

Datasets: ACSIncome ACSEducation Pokec-n Pokec-z Fitzpatrick

Benchmarks: ACSIncome ACSEducation Pokec-n Pokec-z Fitzpatrick

Metrics: efficiency (prediction set size) worst-case fairness disparity coverage

Baselines: Federated Conformal Prediction (FCP) with T-Digest

Builds on: Conformal Fairness Federated Conformal Prediction APS RAPS DAPS FedAvg

Closest papers by meaning

• A Generic Framework for Conformal Fairness (ICLR 2025 · similarity 0.92)
• FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.88)
• FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.87)
• On the Burden of Achieving Fairness in Conformal Prediction (NeurIPS 2026 · similarity 0.86)
• Certifiably Byzantine-Robust Federated Conformal Prediction (ICML 2024 · similarity 0.86)
• FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.86)