Title: Personalized Statistical Validity in Multi-Agent Conformal Prediction

Abstract: Accurate uncertainty quantification is critical for machine learning applications involving high-stakes decision-making. While conformal prediction offers a principled approach to this problem, its effectiveness is often hindered by issues such as data heterogeneity, privacy restrictions, and insufficient local calibration data. Current methods in multi-agent environments fail to adequately resolve these complexities simultaneously; they either provide guarantees that only apply to average performance across agents or sacrifice validity when dealing with heterogeneous data distributions. To address these gaps, we introduce Personalized Federated Weighted Conformal Prediction (PFWCP). This framework integrates local density ratio weighting with weighted quantile aggregation, a combination designed to mitigate heterogeneity while strictly maintaining privacy standards. PFWCP ensures asymptotic validity for both marginal and calibration-conditional coverage for every individual agent involved, and it is compatible with one-shot communication protocols. Our theoretical work derives an adjustment for coverage variance, defined by an effective sample size metric, which is essential for weighted conformal prediction scenarios. Empirical evaluations conducted on both synthetic and real-world datasets demonstrate that our approach achieves superior calibration quality compared to existing state-of-the-art federated conformal baselines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC