Explaining Drift using Shapley Values

TL;DR

The paper tackles the lack of principled root-cause explanations for concept drift by introducing DBShap, a Shapley-based framework that operates on distributions to attribute drift to changes in $P(x)$ (virtual drift) and $P(y|x)$ (real drift). It extends Shapley values to functions over distributions using surrogate binary features and defines a distribution-aware risk decomposition with explicit formulas for the contributions from distributional changes. DBShap combines these theoretical developments with practical approximations (binning, empirical risk, Baseline SHAP) and validates the approach on synthetic and real drift benchmarks, demonstrating meaningful attributions of drift drivers. The method facilitates diagnosing and mitigating drift, enabling more informed decisions about retraining or adapting models in dynamic environments.

Abstract

Machine learning models often deteriorate in their performance when they are used to predict the outcomes over data on which they were not trained. These scenarios can often arise in real world when the distribution of data changes gradually or abruptly due to major events like a pandemic. There have been many attempts in machine learning research to come up with techniques that are resilient to such Concept drifts. However, there is no principled framework to identify the drivers behind the drift in model performance. In this paper, we propose a novel framework - DBShap that uses Shapley values to identify the main contributors of the drift and quantify their respective contributions. The proposed framework not only quantifies the importance of individual features in driving the drift but also includes the change in the underlying relation between the input and output as a possible driver. The explanation provided by DBShap can be used to understand the root cause behind the drift and use it to make the model resilient to the drift.