Prathyush Parvatharaju, Ramesh Doddaiah, Tom Hartvigsen, Elke Rundensteiner

Paper: #insert link

TS-Explain supports command line usage and provides python based API.

pip install tsexp

# Explain a single instance and output to an image
tsexp -i 131 -f data.csv -m xyz.model -o saliency.png

# Explain a dataset
tsexp -a pert -f data.csv -m xyz.model -o saliency.csv

from tsexp import PERT

# Explain a single instance
saliency = PERT.explain_instance(...)

# Explain Dataset
saliencies = PERT.explain(...)

Timeseries Explain

Explainable classification is essential to high-impact settings where practitioners require evidence to support their decisions. However, state-of-the-art deep learning models suffer from a lack of trans- parency in how they derive their predictions. One common form of explainability, termed attribution-based explainability, identi- fies which input features are used by the classifier for its predic- tion. While such explainability for image classifiers has recently received focus, little work has been done to-date to address ex- plainability for deep time series classifiers. In this work, we thus propose PERT, a novel perturbation-based explainability method designed explicitly for time series that can explain any classifier. PERT adaptively learns to perform timestep-specific interpolation to perturb instances and explain a black-box model’s predictions for a given instance, learning which timesteps lead to different be- havior in the classifier’s predictions. For this, PERT pairs two novel complementary techniques into an integrated architecture: a Priori- tized Replacement Selector that learns to select the best replacement time series from the background dataset specific to the instance-of- interest with a novel and learnable Guided-Perturbation Function, that uses the replacement time series to carefully perturb an input instance’s timesteps and discover the impact of each timestep on a black-box classifier’s final prediction. Across our experiments recording three metrics on nine publicly-available datasets, we find that PERT consistently outperforms the state-of-the-art explain- ability methods. We also show a case study using the CricketX dataset that demonstrates PERT succeeds in finding the relevant regions of gesture recognition time series.

Python 3.7+

# Bare installation
git clone https://github.com/kingspp/pert

# With pre-trained models and datasets
git clone --recurse-submodules -j8 https://github.com/kingspp/pert

# Install requirements
cd pert && pip install -r requirements.txt

python3 main.py --pname TEST --task_id10 \
--run_mode turing --jobs_per_task 20 \
--algo pert \
--dataset wafer \
--enable_dist False \
--enable_lr_decay False \
--grad_replacement random_instance \
--eval_replacement class_mean \
--background_data_perc 100 \
--run_eval True \
--enable_seed True \
--w_decay 0.00 \
--bbm dnn \
--max_itr 500

@inproceedings{parvatharaju2021learning,
  author  = {Parvatharaju, Prathyush and Doddaiah, Ramesh and Hartvigsen, Thomas and Rundensteiner, Elke},
  title   = {Learning Saliency Maps for Deep Time Series Classifiers},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year    = 2021,
}