GAM (Global Attribution Mapping)
Global Explanations for Deep Neural Networks
GAM explains the landscape of neural network predictions across subpopulations.
This implementation is based on "Global Explanations for Neural Networks: Mapping the Landscape of Predictions" (AAAI/ACM AIES 2019).
First generate local attributions using your favorite technique, then
>>> gam = GAM(local_path="local_attributions.csv", distance="spearman", k=2) >>> gam.generate() >>> gam.explanations [[('height', .6), ('weight', .3), ('hair color', .1)], [('weight', .9), ('weight', .05), ('hair color', .05)]] >>> gam.subpopulation_sizes [90, 10] >>> gam.subpopulations # global explanation assignment [0, 1, 0, 0,...] >>> gam.plot() # bar chart of feature importance with subpopulation size
Thank you to Paul Zeng for his contribution to the implementation and valuable feedback.
- Branching: master is protected (need one other reviewer to merge)
- Input/Output: csv (columns: features, rows: local/global attribution)
- Underlying data structures:
To run tests:
$ python -m pytest tests/
- assume local attributions are given
- K is a specified parameter
- accompany csv output with a simple visualization showing top 5 clusters and their top 5 features
- optimize k based on silouhette score
- generate local attributions (using appropriate local method)
We welcome Your interest in Capital One’s Open Source Projects (the “Project”). Any Contributor to the Project must accept and sign an Agreement indicating agreement to the license terms below. Except for the license granted in this Agreement to Capital One and to recipients of software distributed by Capital One, You reserve all right, title, and interest in and to Your Contributions; this Agreement does not impact Your rights to use Your own Contributions for any other purpose.
Code of Conduct
This project adheres to the Open Code of Conduct By participating, you are expected to honor this code.