A short tutorial for multivariate time series explanation using tsCaptum

Abstract

tsCaptum is a Python library that enables explainability for time series classification and regression using saliency maps (i.e., attribution-based explanation). It bridges the gap between popular time series frameworks (e.g., aeon, sktime, sklearn) and explanation libraries like Captum. tsCaptum tackles the computational complexity of explaining long time series by employing chunking techniques, significantly reducing the number of model evaluations required. This allows users to easily apply Captum explainers to any univariate or multivariate time series model or pipeline built using the aforementioned frameworks. tsCaptum is readily available on pypi.org and can be installed with a simple ”pip install tsCaptum” command.