Integrated Gradient Cohort Shapley (IGCS) is a calculation method of local feature attribution that apply multilinear extensions which make integrated gradients equivalent to Shapley values in an original space to the space of indicator functions, not to data space. Then it introduces integrated gradients on the indicator space. Our method is based on empirical distribution similar to Cohort Shapley (CS), and can evaluate feature attributions with a linear complexity to the number of features. In addition, IGCS does not require any differentiability on the model nor other assumptions, since line integrals of IGCS are held in the indicator space.
See the paper:
Hama, N., Mase, M., and Owen, A. B. (2022). Model free Shapley values for high dimensional data. arXiv preprint arXiv:2211.08414.
And details of the cohort shapley for the paper:
Mase, M., Owen, A. B., & Seiler, B. (2019). Explaining black box decisions by Shapley cohort refinement. arXiv preprint arXiv:1911.00467.
This library and the example also include the Insertion and Deletion ABC (Area Between the Curves) defined in the paper:
Hama, N., Mase, M., and Owen, A. B. (2022). Deletion and insertion tests in regression models. arXiv preprint arXiv:2205.12423
Install the package locally with pip command.
git clone https://github.com/cohortshapley/cohortintgrad
pip install -e cohortintgradThis code is tested on:
For example notebooks, we need:
See Jupyter notebook examples in example directory.