Timeseries Saliency Maps: Explaining models across multiple domains

Official reproduction repository for the paper "Timeseries Saliency Maps: Explaining models across multiple domains".

Our plug-and-play Tensorflow/PyTorch library for Cross-domain IG can be found here.

Absract

Traditional saliency map methods, popularized in computer vision, highlight individual points (pixels) of the input which contribute the most to the model's output. However, in time-series they offer limited insights as semantically meaningful features are often found in other domains. We introduce Cross-domain Integrated Gradients, a generalization of Integrated Gradients. Our method enables feature attributions on any domain which can be formulated as an invertible, differentiable transformation of the time-domain. Crucially, our derivation extends the original Integrated Gradients into the complex domain, enabling frequency-based attributions. We provide the necessary theoretical guarantees, namely path-independence and completeness. Our approach reveals interpretable, problem-specific attributions that time-domain methods cannot capture, on three real-world tasks — wearable-sensor heart-rate extraction, electroencephalography-based seizure detection, and zero-shot time-series forecasting. We release an open-source Tensorflow/PyTorch library to enable plug-and-play cross-domain explainability for time-series models. These results demonstrate ability of Cross‑domain Integrated Gradients to provide semantically meaningful insights in time‑series models that are impossible with traditional time‑domain saliency.

Experiments

Install requirements

Each sub-project has its own requirements and should be run on different environment to avoid version conflicts. The requirements.txt file for each sub-project is located at the corresponding folder. See the README files in each subdirectory:

1. Preliminary Cross-Domain Integrated Gradients exploration
2. Frequency-Domain IG
3. IG in the Independent Component Analysis domain
4. Seasonal-Trend IG

The code was developed and tested on Python v3.10.16.

Heart Rate Inference

Experiments for heart rate inference from photoplethysmography signals can be found in the ppg_kidppg folder.

Epilepsy detection

Experiments for epilepsy detection from electroencephalography (EEG) signals are located in eeg_zhu_transformer.

The zhu_transformer implementation along with the weights can be found here.

Foundation model forecasting

The experiments for time-series forecasting using a foundation model are found in timesfm.

For the foundation model we are using TimesFM.

Reference

TODO