Explaining Binary Time-Series Classification with Counterfactuals in an Industrial Use Case

Counterfactual explanations are a common approach tomachine learning interpretability as they indicate the differ-ence between two predicted classes. However, automati-cally extracting sample data points to explain the decisionof time-series models is still an open research issue. There-fore, we propose a technique that can automatically identifyfactual and counterfactual examples from within the neigh-borhood of data points requiring explanation. We furtherextract human interpretable time-series features to explainmodel decisions. We apply these techniques to explain thedecisons of a univariate binary time-series classifier trainedfor predicting the health state of a turbofan engine.