[Paper Review] Counterfactual Explanations of Concept Drift

Fabian Hinder, Barbara Hammer|arXiv (Cornell University)|Jun 23, 2020

Data Stream Mining Techniques28 references4 citations

TL;DR

This paper introduces a novel framework for explaining concept drift in machine learning by identifying and tracking characteristic data samples over time using counterfactual explanations. It formalizes drift as changes in representative prototypes and proposes an efficient algorithm that outperforms baselines in identifying drift-relevant features, especially with resampling and weighting, enhancing model interpretability and enabling informed human intervention in lifelong learning systems.

ABSTRACT

The notion of concept drift refers to the phenomenon that the distribution, which is underlying the observed data, changes over time; as a consequence machine learning models may become inaccurate and need adjustment. While there do exist methods to detect concept drift or to adjust models in the presence of observed drift, the question of explaining drift has hardly been considered so far. This problem is of importance, since it enables an inspection of the most prominent features where drift manifests itself; hence it enables human understanding of the necessity of change and it increases acceptance of life-long learning models. In this paper we present a novel technology, which characterizes concept drift in terms of the characteristic change of spatial features represented by typical examples based on counterfactual explanations. We establish a formal definition of this problem, derive an efficient algorithmic solution based on counterfactual explanations, and demonstrate its usefulness in several examples.

Motivation & Objective

• To address the lack of explainable methods for concept drift, which hinders human understanding and intervention in adaptive machine learning systems.
• To formalize concept drift not just as statistical change, but as evolution in representative data samples (prototypes) over time.
• To develop an efficient algorithm that identifies and tracks characteristic data points whose changes most clearly reflect underlying drift.
• To enable human users to inspect and understand *why* model adaptation is necessary by visualizing how key prototypes evolve.
• To evaluate the method’s robustness and effectiveness across synthetic and real-world datasets with known and unknown ground truth.

Proposed method

• Formalizes concept drift as the change in characteristic data points (prototypes) across time, using counterfactual reasoning to explain how and why a sample's classification or behavior shifts.
• Defines two key metrics: identifiability (i) and consistency (C), to evaluate the quality of identified prototypes in capturing drift-relevant changes.
• Applies clustering algorithms (k-means, affinity propagation, mean shift) with resampling and weighting to improve prototype selection, especially in high-dimensional or overlapping data.
• Uses a two-stage process: first, identify candidate prototypes via clustering; second, refine them using counterfactual perturbations to assess their sensitivity to drift.
• Employs resampling and weighting to enhance the stability and representativeness of selected prototypes, particularly in complex or noisy distributions.
• Evaluates performance using both synthetic data with known ground truth and real-world datasets (digits, wine, boston, diabetes, faces), comparing methods across 30 runs.

Experimental results

Research questions

• RQ1 How can concept drift be meaningfully explained through changes in representative data samples rather than just statistical shifts?
• RQ2 Which clustering and refinement strategies best identify prototypes whose evolution most clearly reflects underlying drift?
• RQ3 To what extent does resampling and weighting improve the identifiability and consistency of detected prototypes in drift explanation?
• RQ4 How do different algorithms (k-means, affinity propagation, mean shift) compare in detecting drift-relevant prototypes across diverse data distributions?
• RQ5 Can the proposed method enable human users to intuitively understand *why* a model needs adaptation, based on visualizing prototype evolution?

Key findings

• Affinity propagation and k-means with resampling and weighting consistently outperformed baseline methods in identifying drift-relevant prototypes across all benchmark datasets.
• On the digits and faces datasets, k-means with resampling and weighting achieved mean identifiability scores of 0.84 and 0.80, respectively, significantly outperforming the baseline k-means (0.49 and 0.29).
• Affinity propagation achieved the highest mean identifiability (0.87) on the digits dataset, indicating strong performance in detecting salient drift patterns.
• Mean shift showed high variance and poor performance on regression and high-class datasets (e.g., 0.34 on faces), suggesting instability in prototype selection.
• In synthetic data with known ground truth, resampling significantly improved identifiability (i), while dimensionality had a notable impact on consistency (C), especially in high-dimensional normal distributions.
• The results suggest that optimizing for consistency (C) directly is less effective than focusing on identifiability (i), as the former may be outweighed by density estimation errors.