State-of-the-art methods for explaining predictions fromtime series involve learning an instance-wise saliency mask for each timestep; however, many types of time series are difficult to interpret in thetime domain, due to the inherently complex nature of the data. Instead,we propose to view time series explainability as saliency maps over inter-pretable parts, leaning on established signal processing methodologyon signal decomposition. Specifically, we propose a new method calledFLEXtime that uses a bank of bandpass filters to split the time seriesinto frequency bands. Then, we learn the combination of these bandsthat optimally explains the model’s prediction. Our extensive evaluationshows that, on average, FLEXtime outperforms state-of-the-art explain-ability methods across a range of datasets. FLEXtime fills an importantgap in the current time series explainability methodology and is a valu-able tool for a wide range of time series such as EEG and audio. Codeis available at https://github.com/theabrusch/FLEXtime.

‍