ETD

• agnostic
• autoencoder
• black box
• blackbox
• explainability
• interpretability
• neural network
• shap
• surrogate
• time series

This study presents an agnostic approach to explain the predictions of time series black-box classifiers. Through an autoencoder, this method generates a local neighborhood of the instance to be explained, and then learns a local decision tree classifier to extract rules and counterfactuals. The final explanation is composed by exemplar and counterexemplar time series, and by a shapelet based verbose and graphical decision rule clarifying the black-box decision.