explainy - black-box model explanations for humans

explainy is a library for generating machine learning models explanations in Python. It uses methods from Machine Learning Explainability and provides a standardized API to create feature importance explanations for samples.

The API is inspired by scikit-learn and has three core methods explain(), plot() and, importance(). The explanations are generated in the form of texts and plots.

explainy comes with four different algorithms to create either global or local and contrastive or non-contrastive model explanations.

Description:

• global: explanation of system functionality (all samples have the same explanation)
• local: explanation of decision rationale (each sample has its own explanation)
• contrastive: tracing of decision path (differences to other outcomes are described)
• non-contrastive: parameter weighting (the feature importance is reported)

Documentation

https://explainy.readthedocs.io

Install explainy

pip install explainy

Further, install graphviz (version 9.0.0 or later) for plotting tree surrogate models:

Windows

choco install graphviz

Mac

brew install graphviz

Linux: Ubuntu packages

sudo apt install graphviz

Further details on how to install graphviz can be found in the official graphviz docs.

Also, make sure that the folder with the dot executable is added to your systems PATH. You can find further details here.

Usage

📚 A comprehensive example of the explainy API can be found in this

📖 Or in the example section of the documentation

Initialize and train a scikit-learn model:

import pandas as pd
from sklearn.datasets import load_diabetes
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

diabetes = load_diabetes()
X_train, X_test, y_train, y_test = train_test_split(
    diabetes.data, diabetes.target, random_state=0
)
X_test = pd.DataFrame(X_test, columns=diabetes.feature_names)
y_test = pd.DataFrame(y_test)

model = RandomForestRegressor(random_state=0)
model.fit(X_train, y_train)

Initialize the PermutationExplanation (or any other explanation) object and pass in the trained model and the to be explained dataset.

Addtionally, define the number of features used in the explanation. This allows you to configure the verbosity of your exaplanation.

Set the index of the sample that should be explained.

from explainy.explanations import PermutationExplanation

number_of_features = 4

explainer = PermutationExplanation(
    X_test, y_test, model, number_of_features
)

Call the explain() method and print the explanation for the sample (in case of a local explanation every sample has a different explanation).

explanation = explainer.explain(sample_index=1)
print(explanation)

The RandomForestRegressor used 10 features to produce the predictions. The prediction of this sample was 251.8.

The feature importance was calculated using the Permutation Feature Importance method.

The four features which were most important for the predictions were (from highest to lowest): 'bmi' (0.15), 's5' (0.12), 'bp' (0.03), and 'age' (0.02).

Use the plot() method to create a feature importance plot of that sample.

explainer.plot()

If your prefer, you can also create another type of plot, as for example a boxplot.

explainer.plot(kind='box')

Finally, you can also look at the importance values of the features (in form of a pd.DataFrame).

feature_importance = explainer.importance()
print(feature_importance)

  Feature  Importance
0     bmi        0.15
1      s5        0.12
2      bp        0.03
3     age        0.02
4      s2       -0.00
5     sex       -0.00
6      s3       -0.00
7      s1       -0.01
8      s6       -0.01
9      s4       -0.01

Features

• Algorithms for inspecting black-box machine learning models
• Support for the machine learning frameworks scikit-learn and xgboost
• explainy offers a standardized API with three core methods explain(), plot(), importance()

Other Machine Learning Explainability libraries to watch

• shap: A game theoretic approach to explain the output of any machine learning model
• eli5: A library for debugging/inspecting machine learning classifiers and explaining their predictions
• alibi: Algorithms for explaining machine learning models
• interpret: Fit interpretable models. Explain blackbox machine learning

Source

Molnar, Christoph. "Interpretable machine learning. A Guide for Making Black Box Models Explainable", 2019. https://christophm.github.io/interpretable-ml-book/

Author

Mauro Luzzatto - Maurol