Temporal Rule-Based Counterfactual Explanations for Multivariate Time Series

Abstract

The black-box nature of machine learning models is the main reason impeding their full adoption in decision-making processes. In order to reduce models' opacity and overpass this challenge, major efforts that aim to increase stakeholders' trust and ensure the fairness of decisions are being made by the data mining community under the Explainable Artificial Intelligence (XAI) paradigm. The two main categories of solutions are 1) developing fully transparent algorithms and 2) providing post hoc explanations. However, the literature is rather scarce when it comes to time series data, and even more so in the context of multivariate time series. In this work, we aim to exploit the discriminative power of shapelets and temporal rules in time series mining and capitalize on their inherent interpretability to develop a model-agnostic, temporal rule counterfactual explainer (TeRCE) for multivariate time series datasets. Counterfactual explanations indicate how should the input change such that the decision output changes too. Thus, they can highly increase the interpretability of black-box models. We test TeRCE on five benchmark datasets from the UEA archive and prove that it produces high-quality counterfactuals. Moreover, we show that in addition to being visually and conceptually interpretable, our approach performs better than the state-of-the-art algorithms in terms of proximity, sparsity, and second in terms of plausibility.