Astrapia - A Friendly XAI Explainer Evaluation Framework

Astrapia, derived from the Greek word 'astrapios' meaning a flash of lighting, is an evaluation framework for comparing local model-agnostic post-hoc explainers. The design of Astrapia is based on a few guiding principles:

• Understandable: Astrapia is written with maintainability in mind. The coumminity is encouraged to both read and contribute to the codebase.
• Extendable: Astrapia originated from a research project at the Technical University of Darmstadt. It was built with XAI research in mind. Many components can be extended to build new state-of-the-art systems.
• Customizable: Post-hoc explainers vary wildly from one to another. Astrapia allows for a wide range of different configurations depending on the needs of each individual use case.

Installation

Start by cloning the repository and move to the project folder.

git clone [email protected]:DataManagementLab/Astrapia.git && cd Astrapia

Run the following command to install necessary dependencies. A symbolic link will be built to astrapia allowing you to change the source code without re-installation.

pip install -r requirements.txt

Documentation

Astrapia provides an extensive documentation at astrapia.readthedocs.io.

Quickstart

Currently, we offer two examples: UCI adult dataset and UCI breast cancer dataset. These examples can be found under notebooks/AstrapiaComparatorDemo.ipynb Here we show you how to use Astrapia to compare different explainers using the UCI adult dataset. First, navigate into data/adult/ and run

python setup_adult.py

Files for the datasets will be generated under the corresponding folder. Now load the dataset:

data = dataset.load_csv_data('adult', root_path='../data')

Import the dependencies

import astrapia as xb
from astrapia import explainers, dataset
from astrapia.comparator import ExplainerComparator
from astrapia.visualize_metrics import print_metrics, load_metrics_from_json
import sklearn.ensemble

Then, train a machine learning classifier that you want to explain.

rf = sklearn.ensemble.RandomForestClassifier(n_estimators=50, n_jobs=5)
rf.fit(xb.utils.onehot_encode(data.data, data), data.target.to_numpy().reshape(-1))
pred_fn = lambda x: rf.predict_proba(xb.utils.onehot_encode(x, data))

Prepare post-hoc explainers that you want to compare. Here we chose LIME and Anchors.

ex_lime = explainers.LimeExplainer(data, pred_fn, discretize_continuous=False)
ex_anchors = explainers.AnchorsExplainer(data, pred_fn, 0.9)

Astrapia offers a convenient interface to compare between explainers by instantiating a ExplainerComparator class and appending the explainer to it:

comp = ExplainerComparator()
comp.add_explainer(ex_anchors, 'ANCHORS 0.9')
comp.add_explainer(ex_lime, 'LIME')

Choose an instance or multiple instances to explain:

comp.explain_instances(data.data.iloc[[0]]) # single instance

or

comp.explain_instances(data.data.iloc[[111, 222, 333, 444]]) # multiple instances

Store metric data as json and assert that storing and reloading data does not modify it.

metric_data = comp.get_metric_data()
comp.store_metrics()
assert load_metrics_from_json('metrics.json') == metric_data

To visualize metrics as tables or bar charts:

# show all explainers
print_metrics(metric_data, plot='table', show_metric_with_one_value=True)
print_metrics(metric_data, plot='bar', show_metric_with_one_value=False)

# show single explainer result
print_metrics(metric_data, explainer='ANCHORS 0.9')
print_metrics(metric_data, plot="bar", explainer='LIME')