Dense-Pixel Views for Attribution Techniques on Time Series (DAVOTS)

DAVOTS visualizes raw time series data, activations of a selected layer, and attributions in a dense pixel visualization using image generation for whole time series datasets on trained models.

Through such a visualization, patterns in the data and the explanations can be revealed. For instance, patterns the neural network learns for the classification of samples.

MLVIS 2023 freeze: 076dffd54a1197da9bb88411887b7290fcae81cf

VDS 2024 current WIP.

Data and Model to explore:

Dataset: Ford A (Training: 3601 / Test: 1320)
Model: 3 Conv1D (feature-maps=[5, 10, 50], kernel=3, stride=2), 2 FC (100, 2)
Attribution: IntegratedGradients, ShapleyValueSampling
Activations: 100 FC

Dataset: Ford A (Training: 3601 / Test: 1320)
Model: 3 ResNet Blocks Conv1D (feature-maps=[10, 20, 20], kernel=[5, 3, 1] per Block, stride=1), 2 FC (100, 2)
Attribution: IntegratedGradients, ShapleyValueSampling
Activations: 100 FC

Check the preprocessing directory and preprocessing.py script.

Start the application

You can start the application using:
docker-compose up
and go to localhost to use and explore the application.

Extend with other data

You can rerun the preprocessing script with other data or models.

Development

For further development, you can make use of the yarn (react-scripts) server with:
docker-compose -f docker-compose-dev.yaml up
and go to localhost:3000.

Libraries

Frontend:

• React (https://reactjs.org/)
• React MUI (https://mui.com/)

Backend:

• Python v3.10
• FastAPI (https://fastapi.tiangolo.com/)
• Pillow (https://pillow.readthedocs.io/en/stable/)
• Pytorch (https://pytorch.org/)
• Captum (https://captum.ai/)
• Numpy (https://numpy.org/)
• Scipy (https://scipy.org/)
• Pandas (https://pandas.pydata.org/)

License

Released under MIT License. See the LICENSE file for details.

Reference

@conference{,
 author = {Schlegel, Udo and Keim, Daniel A.},
 booktitle = {Machine Learning Methods in Visualisation for Big Data (MLVIS) 2023 @ EuroVis 2023},
 title = {Interactive dense pixel visualizations for time series and model attribution explanations},
 year = {2023}
}