shaprpy

Python wrapper for the R package shapr.

NOTE: This wrapper is in an EXPERIMENTAL state. Bugs are not unlikely to occur.

Install

The below instructions assume you already have pip and R installed and exposed to the python environment where you want to run shaprpy. Official instructions for installing pip can be found here, and for R here. R can also be installed with pip as follows:

pip install rbase

and conda:

conda install -c r r

Install R-package

The shaprpy Python wrapper requires the development version of the shapr R-package (from the master branch on GitHub). Install it by running the following terminal command from the folder of this readme file (.../shapr/python):

Rscript install_r_packages.R

Install python wrapper

In the folder of this readme file (.../shapr/python), run

pip install -e .

Demo

from sklearn.ensemble import RandomForestRegressor
from shaprpy import explain
from shaprpy.datasets import load_california_housing

dfx_train, dfx_test, dfy_train, dfy_test = load_california_housing()

## Fit model
model = RandomForestRegressor()
model.fit(dfx_train, dfy_train.values.flatten())

## Shapr
explanation = explain(
    model = model,
    x_train = dfx_train,
    x_explain = dfx_test,
    approach = 'empirical',
    phi0 = dfy_train.mean().item(),
)
print(explanation["shapley_values_est"])

shaprpy knows how to explain predictions from models from sklearn, keras and xgboost. For other models, one can provide a custom predict_model function (and optionally a custom get_model_specs) to shaprpy.explain.

See /examples for runnable examples, including an example of a custom PyTorch model.

The /examples/regression_paradigm.py file demonstrates how to use the regression paradigm explained in Olsen et al. (2024). We describe how to specify the regression model, how to enable automatic cross-validation of the model’s hyperparameters, and applying pre-processing steps to the data before fitting the regression models. We refer to Olsen et al. (2024) for when one should use the different paradigms, method classes, and methods.