Ad-hoc explanation for time series classification

In this work, a perturbation-based model-agnostic explanation method for time series classification is presented. One of the main novelties of the proposed method is that the considered perturbations are interpretable and specific for time series. In real-world time series, variations in the speed or the scale of a particular action, for instance, may determine the class, so modifying this type of characteristic leads to ad-hoc explanations for time series. To this end, four perturbations or transformations are proposed: warp, scale, noise, and slice. Given a transformation, an interval of a series is considered relevant for the prediction of a classifier if a transformation in this interval changes the prediction. Another novelty is that the method provides a two-level explanation: a high-level explanation, where the robustness of the prediction with respect to a particular transformation is measured, and a low-level explanation, where the relevance of each region of the time series in the prediction is visualized. In order to analyze and validate our proposal, first some illustrative examples are provided, and then a thorough quantitative evaluation is carried out using a specifically designed evaluation procedure.