Generalized Shapley Additive Explanations

Generalized Shapley Additive Explanations (G-SHAP) is a technique in explainable AI for answering broad questions in machine learning.

Applications

This is just a small sample of the questions G-SHAP can answer.

General classification and regression

Suppose we have a black-box model which diagnoses patients with COVID-19, the flu, or a common cold based on their symptoms. Existing explanatory methods can tell us why our model diagnosed a patient with COVID-19. G-SHAP can answer broader questions, such as how do the symptoms which distinguish COVID-19 from the flu differ from those which distinguish COVID-19 from the common cold?.

Full analysis here.

Intergroup differences

Suppose we have a black-box model which predicts a criminal’s risk of recidivism to determine whether they are eligible for parole. Existing explanatory methods can tell us why our model predicted that a criminal has a high recidivism risk. G-SHAP can answer broader questions, such as why does our model predict that Black criminals have higher recidivism rates than White criminals?.

Full analysis here.

Model performance and failure

Suppose we have a black-box model which forecasts GDP growth based on macroeconomic variables. Existing explanatory methods can tell us why our model forecast 3% GDP growth in a given year. G-SHAP can answer broader questions, such as why did our model fail to forecast the 2008-2009 financial crisis?.

Full analysis here.

Installation

$ pip install gshap

Quickstart

Here we train a support vector classifier to predict whether a criminal will recidivate within two years of release from prison. We use G-SHAP to ask why our model predicts that Black criminals are more likely to recidivate than non-Black criminals.

import gshap
from gshap.datasets import load_recidivism
from gshap.intergroup import IntergroupDifference

from sklearn.svm import SVC

recidivism = load_recidivism()
X, y = recidivism.data, recidivism.target
clf = SVC().fit(X, y)

g = IntergroupDifference(group=X['black'], distance='relative_mean_distance')
explainer = gshap.KernelExplainer(clf.predict, X, g)
explainer.gshap_values(X, nsamples=10)

Out:

array([ 0.01335252,  0.24884556,  0.00132373, -0.0025238 , -0.00151837,
    0.40453822,  0.01636782,  0.07666043, -0.00056414,  0.00966583])

The sum of the G-SHAP values is the relative difference in predicted recidivism rates. The model predicts that Black criminals are 75% more likely to recidivate.

The variables most responsible for this difference are number of prior convictions (index 5; 40%), age (index 1; 25%), and race (index 7; 8%).

Citation

@software{bowen2020gshap,
  author = {Dillon Bowen},
  title = {Generalized Shapley Additive Explanations},
  url = {https://dsbowen.github.io/gshap/},
  date = {2020-05-19},
}

License

Users must cite G-SHAP in any publications which use this software.

G-SHAP is licensed with the MIT License.