Counterfactual Explanation Bake-Off: A Review and Experimental Evaluation for Time Series Classification

Abstract Machine learning models often excel in prediction tasks but frequently lack interpretability, limiting their use in critical domains where understanding and trust are essential. Counterfactual explanations bridge this gap by identifying the minimal feature modifications necessary to alter a model’s prediction to a desired outcome. This paper offers a comprehensive review and evaluation of counterfactual explanation methods specifically for univariate time series classification. By categorizing methods based on their underlying principles, we highlight their strengths, limitations, and the inherent compromises between properties such as validity, proximity, sparsity, and plausibility. Our experimental evaluation assesses these key properties across diverse datasets. The results make it evident that no current method achieves all these properties simultaneously. By synthesizing advancements and pinpointing areas for improvement, this work aims to guide future research and foster the development of interpretable AI solutions for time series applications.