shapwrap

A wrapper for easier and slightly more extended SHAP library usage. Just makes this easier to plug in and not worry about formatting of the package.

Extended features:

• group categorical features effect via summary_group. Super useful if you use pd.get_dummies() in your pipelines, so a single feature effect gets obfuscated behind multiple dummies. This will try to group it back together by common prefix.
• feature_breakout will calculate mean Shap contribution from a particular feature's value. The idea is to get a sense of which feature-value pairs are driving most of the model decisions. For continuous feature it will split it into 3 quantiles.

Setup

Install via:

pip install shapwrap

Or clone this package and install as editable install:

pip install -e .

Followed this for pip package setup: https://betterscientificsoftware.github.io/python-for-hpc/tutorials/python-pypi-packaging/#what-is-pip

How to use

import pandas as pd
from sklearn.datasets import load_wine
from shapwrap.wrapper import ShapExplanation
from sklearn.linear_model import LinearRegression

data = load_wine(as_frame=True)
X, y = data.data, data.target

clf = LinearRegression()
clf = clf.fit(X, y)

expl = ShapExplanation(data=X, model = clf)

expl.plot(plot_type='summary', save_plot_path='summary.png')
expl.plot(plot_type='summary_group')
expl.plot(plot_type='decision_paths')
expl.plot(plot_type='feature_breakout')
expl.plot(plot_type='dependence')