Structured Credal Learning

2026-03-14

AI Summary

Real-world learning tasks are often simultaneously affected by covariate shift and label noise, yet existing methods struggle to disentangle these two sources of uncertainty. This work proposes a Structured Trustworthy Learning framework that explicitly decouples covariate shift from label noise for the first time, revealing a gating effect of covariates on label disagreement and eliminating reliance on heuristic robustness parameters. By constructing structured trustworthy sets based on geometric bounds derived from the total variation diameter, and integrating finite-sample concentration inequalities with discrete minimax optimization, the method enables fine-grained modeling of composite uncertainties. The approach offers theoretical guarantees under fixed covariate settings, admits efficient computation, and significantly enhances robust learning performance.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Reasoning under Uncertainty: Stochastic Optimization
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Real-world learning tasks often encounter uncertainty due to covariate shift and noisy or inconsistent labels. However, existing robust learning methods merge these effects into a single distributional uncertainty set. In this work, we introduce a novel structured credal learning framework that explicitly separates these two sources. Specifically, we derive geometric bounds on the total variation diameter of structured credal sets and demonstrate how this quantity decomposes into contributions from covariate shift and expected label disagreement. This decomposition reveals a gating effect: covariate modulates how much label disagreement contributes to the joint uncertainty, such that seemingly benign covariate shifts can substantially increase the effective uncertainty. We also establish finite-sample concentration bounds in a fixed covariate regime and demonstrate that this quantity can be efficiently estimated. Lastly, we show that robust optimization over these structured credal sets reduces to a tractable discrete min-max problem, avoiding ad-hoc robustness parameters. Overall, our approach provides a principled and practical foundation for robust learning under combined covariate and label mechanism ambiguity.

Problem

• covariate shift
• label noise
• distributional uncertainty
• robust learning
• credal sets

Innovation

• structured credal learning
• covariate shift
• label disagreement
• total variation diameter
• min-max optimization