Contrastive Explanation (Foil Trees)

Contrastive and counterfactual explanations for machine learning (ML)

Marcel Robeer (2018-2020), TNO/Utrecht University

Contents

1. Introduction
2. Publications: citing this package
3. Example usage
4. Documentation: choices for problem explanation
5. License

Introduction

Contrastive Explanation provides an explanation for why an instance had the current outcome (fact) rather than a targeted outcome of interest (foil). These counterfactual explanations limit the explanation to the features relevant in distinguishing fact from foil, thereby disregarding irrelevant features. The idea of contrastive explanations is captured in this Python package ContrastiveExplanation. Example facts and foils are:

Publications

One scientific paper was published on Contrastive Explanation / Foil Trees:

• J. van der Waa, M. Robeer, J. van Diggelen, M. Brinkhuis, and M. Neerincx, "Contrastive Explanations with Local Foil Trees", in 2018 Workshop on Human Interpretability in Machine Learning (WHI 2018), 2018, pp. 41-47. [Online]. Available: http://arxiv.org/abs/1806.07470

It was developed as part of a Master's Thesis at Utrecht University / TNO:

• M. Robeer, "Contrastive Explanation for Machine Learning", MSc Thesis, Utrecht University, 2018. [Online]. Available: https://dspace.library.uu.nl/handle/1874/368081

Citing this package

@inproceedings{vanderwaa2018,
  title={{Contrastive Explanations with Local Foil Trees}},
  author={van der Waa, Jasper and Robeer, Marcel and van Diggelen, Jurriaan and Brinkhuis, Matthieu and Neerincx, Mark},
  booktitle={2018 Workshop on Human Interpretability in Machine Learning (WHI)},
  year={2018}
}

Example usage

As a simple example, let us explain a Random Forest classifier that determine the type of flower in the well-known Iris flower classification problem. The data set comprises 150 instances, each one of three types of flowers (setosa, versicolor and virginica). For each instance, the data set includes four features (sepal length, sepal width, petal length, petal width) and the goal is to determine which type of flower (class) each instance is.

Steps

First, train the 'black-box' model to explain

from sklearn import datasets, model_selection, ensemble
seed = 1

# Train black-box model on Iris data
data = datasets.load_iris()
train, test, y_train, y_test = model_selection.train_test_split(data.data, 
                                                                data.target, 
                                                                train_size=0.80, 
                                                                random_state=seed)
model = ensemble.RandomForestClassifier(random_state=seed)
model.fit(train, y_train)

Next, perform contrastive explanation on the first test instance (test[0]) by wrapping the tabular data in a DomainMapper, and then using method ContrastiveExplanation.explain_instance_domain()

# Contrastive explanation
import contrastive_explanation as ce

dm = ce.domain_mappers.DomainMapperTabular(train, 
                                           feature_names=data.feature_names,
					   contrast_names=data.target_names)
exp = ce.ContrastiveExplanation(dm, verbose=True)

sample = test[0]
exp.explain_instance_domain(model.predict_proba, sample)

[OUT] "The model predicted 'setosa' instead of 'versicolor' because 'sepal length (cm) <= 6.517 and petal width (cm) <= 0.868'"

The predicted class using the RandomForestClassifier was 'setosa', while the second most probable class 'versicolor' may have been expected instead. The difference of why the current instance was classified 'setosa' is because its sepal length is less than or equal to 6.517 centimers and its petal width is less than or equal to 0.868 centimers. In other words, if the instance would keep all feature values the same, but change its sepal width to more than 6.517 centimers and its petal width to more than 0.868 centimers, the black-box classifier would have changed the outcome to 'versicolor'.

More examples

For more examples, check out the attached Notebook.

Documentation

Several choices can be made to tailor the explanation to your type of explanation problem.

Choices for problem explanation

FactFoil

Used for determining the current outcome (fact) and the outcome of interest (foil), based on a foil_method (e.g. second most probable class, random class, greater than the current outcome). Foils can also be manually selected by using the foil=... optional argument of the ContrastiveExplanation.explain_instance_domain() method.

Explanators

Method for forming the explanation, either using a Foil Tree (TreeExplanator) as described in the paper, or using a prototype (PointExplanator, not fully implemented). As multiple explanations hold, one can choose the foil_strategy as either 'closest' (shortest explanation), 'size' (move the current outcome to the area containing most samples of the foil outcome), 'impurity' (most informative foil area), or 'random' (random foil area)

Domain Mappers

For handling the different types of data:

• Tabular (rows and columns)
• Images (rudimentary support)

Maps to a general format that the explanator can form the explanation in, and then maps the explanation back into this format. Ensures meaningful feature names.

License

ContrastiveExplanation is BSD-3 Licensed.