Dynamask - Explaining Time Series Predictions with Dynamic Masks

Code Author: Jonathan Crabbé (jc2133@cam.ac.uk)

This repository contains the implementation of Dynamask, a method to identify the features that are salient for a model to issue its prediction when the data is represented in terms of time series. For more details on the theoretical side, please read our ICML 2021 paper: 'Explaining Time Series Predictions with Dynamic Masks'.

Part of the experiments in our paper are relying on FIT, another repository associated to the NeurIPS 2021 paper : 'What went wrong and when? Instance-wise feature importance for time-series black-box models'. We have included all the relevant files in the folder fit.

Installation

To install the relevant packages from shell:

Clone the repository
Create a new virtual environment with Python 3.8

Run the following command from the repository folder:

pip install -r requirements.txt #install requirements

When the packages are installed, Dynamask can directly be used.

Toy example

It is very easy to fit a mask on a time series model. Bellow, you can find a toy demonstration where we fit a mask to an input time series. In this case, the mask area is fixed to 0.1 (the 10% most important features are highlighted by the mask). All the relevant code can be found in the file mask.

import torch
from attribution.mask import Mask
from attribution.perturbation import GaussianBlur
from utils.losses import mse

torch.manual_seed(42)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Define a pseudo-black box:
def black_box(input):
   output = input[-1, :]  # The black-box returns the features of the last time step
   return output
# Define a random input:
X = torch.randn(10, 3).to(device) # The shape of the input has to be (T, N_features)

# Fit a mask to the input with a Gaussian Blur perturbation:
pert = GaussianBlur(device)
mask = Mask(pert, device)
mask.fit(X, black_box, loss_function=mse, keep_ratio=0.1, size_reg_factor_init=0.01) # Select the 10% most important features

# Plot the resulting saliency map:
mask.plot_mask()

If the proportion of features to select is unkown, a good approach is to fit a group of masks with different areas. Then, the extremal mask can be extracted from the group. The relevant code can be found in the file mask_group.

import torch
from attribution.mask_group import MaskGroup
from attribution.perturbation import GaussianBlur
from utils.losses import mse

torch.manual_seed(42)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Define a pseudo-black box:
def black_box(input):
   output = input[-1, :]  # The black-box returns the features of the last time step
   return output

# Define a random input:
X = torch.randn(10, 3).to(device) # The shape of the input has to be (T, N_features)

# Fit a group of masks to the input with a Gaussian Blur perturbation:
areas = [.1, .15, .2, .25] # These are the areas of the different masks
pert = GaussianBlur(device)
masks = MaskGroup(pert, device)
masks.fit(X, black_box, loss_function=mse, area_list=areas, size_reg_factor_init=0.01)

# Extract the extremal mask:
epsilon = 0.01
mask = masks.get_extremal_mask(threshold=epsilon)

# Plot the resulting saliency map:
mask.plot_mask()

Replicate experiments

All experiments in the ICML paper can be replicated easily. The necessary code is in experiments. Bellow , we detail the procedure for each experiment.

Replicate the Rare experiments

Run the following command from the repository folder:

python -m experiments.rare_feature # Runs the Rare Feature experiment
python -m experiments.rare_time # Runs the Rare Time experiment

To do the experiment with various seeds, please add the following specification to these commands:

Options:
--cv # An integer that sets the random seed (first run cv=0 , second run cv=1, ...)

The results of these experiments are saved in the two following folders: Rare Feature and Rare Time. To process the results and compute the associated metrics run:

python -m experiments.results.rare_feature.get_results
python -m experiments.results.rare_time.get_results

The following options need to be specified:

Options:
--CV # The number of runs you have done for the experiment
--explainers # The baselines you have used among: dynamask, fo, fp, ig, shap (separated by a space)

Replicate the State experiment

Run this command to generate the synthetic data and store it in data/state:

python -m fit.data_generator.state_data --signal_len 200 --signal_num 1000

Run the following command to fit a model together with a baseline saliency method:

python -m fit.evaluation.baselines --explainer fit --train

To do the experiment with various baselines, please change the explainer:

Options:
--explainer # The baselines can be: fit, lime, retain, integrated_gradient, deep_lift, fo, afo, gradient_shap
--train # Only put this option when fitting the FIRST baseline (this is to avoid retraining a model for each baseline)
--cv # An integer that sets the random seed (first run cv=0 , second run cv=1, ...)

The models and baselines saliency maps are all saved in this folder. Now fit a mask for each of these time series by running:
```
python -m experiments.state
```
Please use the same --cv option as for the previous command.

The masks are all saved in this folder. To process the results and compute the associated metrics run:

python -m experiments.results.state.get_results

The following options need to be specified:

Options:
--CV # The number of runs you have done for the experiment
--explainers # The baselines you have used among: dynamask, fo, afo, deep_lift, fit, gradient_shap, integrated_gradient, lime, retain (separated by a space)

Replicate the MIMIC experiment

MIMIC-III is a private dataset. For the following, you need to have You need to have the MIMIC-III database running on a local server. For more information, please refer to the official MIMIC-III documentation.
Run this command to acquire the data and store it:
```
python fit/data_generator/icu_mortality.py --sqluser YOUR_USER --sqlpass YOUR_PASSWORD
```
If everything happens properly, two files named adult_icu_vital.gz and adult_icu_lab.gz are stored in data/mimic.
Run this command to preprocess the data:
```
python fit/data_generator/data_preprocess.py
```
If everything happens properly, a file patient_vital_preprocessed.pkl is stored in data/mimic.

Run the following command to fit a model together with a baseline saliency method:

python -m fit.evaluation.baselines --data mimic --explainer fit --train

To do the experiment with various baselines, please change the explainer:

Options:
--explainer # The baselines can be: fit, lime, retain, integrated_gradient, deep_lift, fo, afo, gradient_shap
--train # Only put this option when fitting the FIRST baseline (this is to avoid retraining a model for each baseline)
--cv # An integer that sets the random seed (first run cv=0 , second run cv=1, ...)

The models and baselines saliency maps are all saved in this folder. Now fit a mask for each of these time series by running:
```
python -m experiments.mimic
```
Please use the same --cv option as for the previous command.
```
Options:
--cv # Same as in the previous command
--area # The area of the mask to fit (a number between 0 and 1)
```

The masks are all saved in this folder. To process the results and compute the associated metrics run:

python -m experiments.results.state.plot_benchmarks

The following options need to be specified:

Options:
--CV # The number of runs you have done for the experiment
--explainers # The baselines you have used among: dynamask, fo, afo, deep_lift, fit, gradient_shap, integrated_gradient, lime, retain (separated by a space)
--areas # The mask areas that you have computed (separated by a space)

The resulting plots are saved in this folder.

Citing

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

@inproceedings{Crabbe2021Dynamask,
  title={Explaining Time Series Predictions with Dynamic Masks},
  author={Crabbé, Jonathan and van der Schaar, Mihaela},
  year={2021},
  booktitle={Proceedings of the 38-th International Conference on Machine Learning (ICML 2021)},
  organization={PMLR}
}

L-forecaster/Dynamask

Dynamask - Explaining Time Series Predictions with Dynamic Masks

Installation

Toy example

Replicate experiments

Replicate the Rare experiments

Replicate the State experiment

Replicate the MIMIC experiment

Citing