Alibi is an open source Python library aimed at machine learning model inspection and interpretation. The initial focus on the library is on black-box, instance based model explanations.

• Documentation

If you're interested in outlier detection, concept drift or adversarial instance detection, check out our sister project alibi-detect.

• Provide high quality reference implementations of black-box ML model explanation and interpretation algorithms
• Define a consistent API for interpretable ML methods
• Support multiple use cases (e.g. tabular, text and image data classification, regression)

Alibi can be installed from PyPI:

pip install alibi

This will install alibi with all its dependencies:

  beautifulsoup4
  numpy
  Pillow
  pandas
  requests
  scikit-learn
  spacy
  scikit-image
  tensorflow

To run all the example notebooks, you may additionally run pip install alibi[examples] which will install the following:

  seaborn
  Keras

These algorithms provide instance-specific (sometimes also called local) explanations of ML model predictions. Given a single instance and a model prediction they aim to answer the question "Why did my model make this prediction?" The following algorithms all work with black-box models meaning that the only requirement is to have acces to a prediction function (which could be an API endpoint for a model in production).

The following table summarizes the capabilities of the current algorithms:

• Anchor explanations (Ribeiro et al., 2018)

  • Documentation
  • Examples: income prediction, Iris dataset, movie sentiment classification, ImageNet, fashion MNIST

• Contrastive Explanation Method (CEM, Dhurandhar et al., 2018)

  • Documentation
  • Examples: MNIST, Iris dataset

• Counterfactual Explanations (extension of Wachter et al., 2017)

  • Documentation
  • Examples: MNIST

• Counterfactual Explanations Guided by Prototypes (Van Looveren et al., 2019)

  • Documentation
  • Examples: MNIST, Boston housing dataset, Adult income (one-hot), Adult income (ordinal)

These algorihtms provide instance-specific scores measuring the model confidence for making a particular prediction.

(1) Depending on model

(2) May require dimensionality reduction

• Trust Scores (Jiang et al., 2018)
  • Documentation
  • Examples: MNIST, Iris dataset
• Linearity Measure
  • Examples: Iris dataset, fashion MNIST

Anchor method applied to the InceptionV3 model trained on ImageNet:

Prediction: Persian Cat	Anchor explanation

Contrastive Explanation method applied to a CNN trained on MNIST:

Prediction: 4	Pertinent Negative: 9	Pertinent Positive: 4

Trust scores applied to a softmax classifier trained on MNIST: