Skip to content

GAM (Global Attribution Mapping) explains the landscape of neural network predictions across subpopulations

License

Notifications You must be signed in to change notification settings

capitalone/global-attribution-mapping

Repository files navigation

GAM (Global Attribution Mapping)

Global Explanations for Deep Neural Networks

GAM explains the landscape of model predictions across subpopulations. This implementation is based on "Global Explanations for Neural Networks: Mapping the Landscape of Predictions" (AAAI/ACM AIES 2019). GAM is a method for clustering any set of local attributions - it is agnostic to the choice of model architecture and XAI method.

Installation

python3 -m pip install gam

Get Started

First generate local attributions using your favorite technique, then:

>>> from gam.gam import GAM
>>> # for a quick example use `attributions_path="tests/test_attributes.csv"`
>>> # Input/Output: csv (columns: features, rows: local/global attribution)
>>> gam = GAM(attributions_path="<path_to_your_attributes>.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

Tests

To run tests:

$ python -m pytest tests/

Contributors

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.

Sign the Individual Agreement

Sign the Corporate Agreement

Code of Conduct

This project adheres to the Open Code of Conduct By participating, you are expected to honor this code.