ES-Mask: Evolutionary Strip Mask for Explaining Time Series Prediction (Student Abstract)

Machine learning models are increasingly used in time series prediction with promising results. The model explanation of falls behind the development and makes less sense to users understanding decisions. This paper proposes ES-Mask, a post-hoc model-agnostic evolutionary strip mask-based saliency approach for applications. ES-Mask designs mask consisting strips same salient value consecutive steps produce binary sustained feature importance scores over easy interpretation series. uses an algorithm search optimal by manipulating rounds, thus is agnostic involving no internal states search. initial experiments on MIMIC-III data set show that outperforms state-of-the-art methods.