XAI for time series classification: evaluating the benefits of model inspection for end-users

Brigt Håvardstun, Cèsar Ferri, Kristian Flikka, Jan Arne Telle

Abstract: We present an XAI tool for time series classification providing model-agnostic instance-based post-hoc explanations, by means of prototypes and counterfactuals. Additionally, our tool allows for model inspection on instances generated by the user, to navigate the boundary between classes. This will allow the user to test and improve their hypotheses when formulating a model of the black box classification. We perform a human-grounded evaluation with forward simulation, to contribute a quantitative end-user evaluation to the field of XAI for time series.