Drift Localization using Conformal Predictions

arXiv stat.ML / 4/22/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper addresses concept drift, focusing on drift localization, i.e., identifying which samples are affected when the underlying data distribution changes over time.
It argues that many existing drift-localization methods rely on local testing schemes that can break down in high-dimensional, low-signal scenarios.
The authors propose a fundamentally different approach using conformal predictions to improve robustness in these challenging settings.
The work evaluates the method on state-of-the-art image datasets and reports performance improvements over common approaches, illustrating the practical value of the conformal-prediction framework.

Abstract

Concept drift -- the change of the distribution over time -- poses significant challenges for learning systems and is of central interest for monitoring. Understanding drift is thus paramount, and drift localization -- determining which samples are affected by the drift -- is essential. While several approaches exist, most rely on local testing schemes, which tend to fail in high-dimensional, low-signal settings. In this work, we consider a fundamentally different approach based on conformal predictions. We discuss and show the shortcomings of common approaches and demonstrate the performance of our approach on state-of-the-art image datasets.