On the Mining of Time Series Data Counterfactual Explanations using Barycenters

Abstract

EXplainable Artificial Intelligence (XAI) methods are increasingly accepted as effective tools to trace complex machine learning models' decision-making processes. There are two underlying XAI paradigms: (1) traditional factual methods and (2) emerging counterfactual models. The first family of methods uses feature attribution techniques that alter the feature space and observe the impact on the decision function. Counterfactual models aim at providing the smallest possible change to the feature vector that can change the prediction outcome. In this paper, we propose TimeX, a new model-agnostic time series counterfactual explanation algorithm that provides sparse, interpretable, and contiguous explanations. We validate our model using real-world time series datasets and show that our approach can generate explanations with up to 20% fewer outliers in comparison with other state-of-the-art competing baselines.