Time Series Explanation Spaces

This is the open source repository for Explanation Space: A New Perspective into Time Series Interpretability and On the Necessity of Multi-Domain Explanation: An Uncertainty Principle Approach for Deep Time Series Models both published in ICDM 2025.

Library Versions

• Python 3.10.9
• Pytorch 2.2.1
• Captum 0.7.0
• tsai 0.3.9
• tslearn 0.6.3

Train a Target Model

Using the following command, you can trained a ResNet model on GunPoint dataset. You can choose any dataset from UCR repository available in tslearn package.

$ python train.py -m ResNet -d GunPoint

To generate an explanation with DeepLIFT, use the following command

$ python Explain.py -m ResNet -d GunPoint -a DeepLift -x Time

You can change the explanation space with -x option. Current options include 'Time', 'Freq', 'TimeFreq', 'Diff', 'MinZero', and 'Diff_back_to_Time'.

$ python Explain.py -m ResNet -d GunPoint -a DeepLift -x Freq

To check if the uncertainty principle is violated you can set the space variable as 'Uncertainty_principle_test'.

$ python Explain.py -m ResNet -d GunPoint -a DeepLift -x Uncertainty_principle_test