marcotcr / anchor

Code for "High-Precision Model-Agnostic Explanations" paper
BSD 2-Clause "Simplified" License
799 stars 114 forks source link

Anchor

This repository has code for the paper High-Precision Model-Agnostic Explanations.

An anchor explanation is a rule that sufficiently “anchors” the prediction locally – such that changes to the rest of the feature values of the instance do not matter. In other words, for instances on which the anchor holds, the prediction is (almost) always the same.

At the moment, we support explaining individual predictions for text classifiers or classifiers that act on tables (numpy arrays of numerical or categorical data). If there is enough interest, I can include code and examples for images.

The anchor method is able to explain any black box classifier, with two or more classes. All we require is that the classifier implements a function that takes in raw text or a numpy array and outputs a prediction (integer)

Installation

The Anchor package is on pypi. Simply run:

pip install anchor-exp

Or clone the repository and run:

python setup.py install

If you want to use AnchorTextExplainer, you have to run the following:

python -m spacy download en_core_web_lg

And if you want to use BERT to perturb inputs (recommended), also install transformers:

pip install torch transformers spacy && python -m spacy download en_core_web_sm

Examples

See notebooks folder for tutorials. Note that from version 0.0.1.0, it only works on python 3.

Citation

Here is the bibtex if you want to cite this work.