Explaining Time Series Classifications Counterfactually: A Human-Centered Evaluation
TL;DR
This work proposes a novel methodology that advances carefully-controlled tests of counterfactual explanations for time series classification and results indicate that counterfactual explanations elicit improved domain understanding and higher confidence in people in comparison to factual explanations.
Abstract
Counterfactual explanations can promote accountability, fairness, and transparency in AI-supported decision-making by demonstrating how changes in input data can alter outcomes, offering actionable feedback for users to revise unfavorable decisions. However, the majority of counterfactual algorithms focus on tabular and image data, ignoring many high-stakes scenarios that rely on time series data. Furthermore, user studies of counterfactuals for time series are exceedingly rare. We address this research gap by proposing a novel methodology that advances carefully-controlled tests of counterfactual explanations for time series classification. In two experiments, participants were asked to classify time series plots and provide confidence ratings, before receiving feedback on the correct classification along with an explanation. The studies manipulated the explanation strategy used and the visualization used. Results indicate that counterfactual explanations elicit improved domain understanding and higher confidence in people in comparison to factual explanations. Wider implications for explanation in time series domains are discussed.
