FLEXtime: Filterbank learning to explain time series

This repository implements FLEXtime (Filterbank Learning to EXplain time series). If you have any questions, please contact me on theb@dtu.dk.

Link to paper: https://arxiv.org/abs/2411.05841.

Downloading and preprocessing the data

AudioMNIST

Follow https://github.com/soerenab/AudioMNIST to download and preprocess the data.

SleepEDF

Initially, download the Sleep Cassette data (https://www.physionet.org/content/sleep-edfx/1.0.0/). Run the following lin with correct root paths and desired output folders to preprocess the data to follow the correct format:

python3 preprocess_sleepedf.py --root_folder /path/to/sleep-casette/ --out_folder /path/to/sleepedf/

PAM, ECG and Epilepsy

Download the datasets from the TimeX experiments: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/B0DEQJ.

Training the models

To train the models for AudioMNIST, SleepEDF and the synthetic dataset, use the scripts in the train_models/ folder. To train the remaining models, use the code provided by the TimeX authors: https://github.com/mims-harvard/TimeX.

Running the experiments

The experiments can now be run using the main.py script. This has to be run once for every dataset and every seed to reproduce the results from the paper. The script saves pickle files containing the computed explanations for each dataset and seeds and runs the faithfulness tests. The computation of the complexity and smoothness scores then happens posthoc in the collect_results.py script.

Citation

If you use this code, please cite the original paper:

@article{brusch2024flextime,
  title={FLEXtime: Filterbank learning to explain time series},
  author={Br{\"u}sch, Thea and Wickstr{\o}m, Kristoffer K and Schmidt, Mikkel N and Jenssen, Robert and Alstr{\o}m, Tommy S},
  journal={arXiv preprint arXiv:2411.05841},
  year={2024}
}