trustyai-explainability / trustyai-explainability-python

Python bindings for TrustyAI's explainability library

TrustyAI Explainability

This repo is the main hub for TrustyAI, containing the core Java library as well as various modules to support the TrustyAI Service, TrustyAI Operator, and TrustyAI Python Library.

Overview

TrustyAI is, at its core, a Java library and service for Explainable AI (XAI). TrustyAI offers fairness metrics, explainable AI algorithms, and various other XAI tools at a library-level as well as a containerized service and Kubernetes deployment.

Directory

• explainability-core, the core TrustyAI Java module, containing fairness metrics, AI explainers, and other XAI utilities.
• explainability-service, TrustyAI-as-a-service, a REST service for fairness metrics and explainability algorithms including ModelMesh integration.
• explainability-arrow, a Java module to facilitate the communication between TrustyAI-Java and TrustyAI-Python using Arrow.
• explainability-connectors, A Java module to interface with different black-box predictive model deployments or inference services. Includes support for KServe / ModelMesh via gRPC.
• explainability-integrationtests A set of integration tests for integrations within Kogito, namely DMN, PMML, and OpenNLP models.

Roadmap

GitHub project page

For More Information

Our preprint TrustyAI Explainability Toolkit is a great source of knowledge of what the core library can offer.

Furthermore, you can reach the dev team on:

• ODH Community Slack
• Zulip
• or by coming to one of our community meetings

Building and Contributing

All contributions are welcome! Before you start please read the contribution guide.

trustyai-explainability-python's Issues

Automate version increase on release pipeline

Fix ReadTheDocs documentation generation

Remove Java async calls from Python calls

Python TrustyAI does not build with Python 3.12

Add TSSaliency to Python TrustyAI

Implement batching

Support TSICE explanations as dataframe

Add TSICE to Tyrus

Dict to map implicit conversion

Automatic visualisation for explainers

Make it possible to score (only) malicious tokens

As detailed in trustyai-explainability/trustyai-explainability#289 (comment) , scoring "malicious" portions of text would help us integrate our MaRCo implementation together with other techniques for rephrasing.

Replace trustyai factory methods with type inference

Hide CompletableFuture API details

Replace setuptools with hatch

Meta license field is wrong on automated PyPi releases

Example: https://test.pypi.org/project/trustyai/0.2.5/

Convert TSICE saliencies to TrustyAI format

Add proper score to CF XGBoost model

Provide 3.8 and 3.9 python versions of TrustyAI image

In ODH/RHODS it might be reasonable to have (same as for other images) a TrustyAI image for python 3.9 (n) as well as python 3.8 (n-1) versions, this probably can be beneficial for easier adoption, depending on user needs.

Create default domain

https://github.com/trustyai-python/module/blob/5014edebeed5bfd00ed006d2dfaa43fab66ccb42/tests/test_counterfactualexplainer.py#L38

Fix TSSaliency plots

TSSaliency plots not displaying the saliencies correctly.

Add script for local dev build

Add Python accessors to feature importances

Expand Python version CI testing matrix

Containerise testing

Move examples to its own repository

Jupyter notebooks and examples in general should moved to its own repo: https://github.com/trustyai-python/examples

Support TSICE explanations as plots

Make implicit conversions global

Remove java.util.List explicit dependency from PredictionProvider

Add external dependencies as optional dependencies

Fix LIME tests for v1.8

Test are written for v1.5 at the moment and won't work with v1.8.

Numpy arrays to Vector features conversion

Update README.md

Move Maven hardcoded repos to properties

Automatically install required JARs

Enable fairness metrics

Fairness metrics and Dataset API should be enabled in python bindings too (see https://issues.redhat.com/browse/FAI-714).

Create implicit conversion for the Output class

Add counterfactuals introductory notebook

Move PyPi publishing to Trusted Providers

With the introduction of "Trusted Publishers" by PyPi, the Python TrustyAI release process should move to use it and replace token based releases.

Add TSLIME to Python TrustyAI

Implicit conversions for all collections

Update py-arrow dependencies

Python TrustyAI uses 7.0.0, latest is 12.0.0.

Move counterfactual configuration to explainer

Upgrade TrustyAI notebook images to use version 0.3.0

Upgrade TrustyAI images to use TrustyAI 0.3.0 version

Automate version sync between bindings and TrustyAI

Add confidence threshold example to CF notebook

Move SimplePrediction to models package

SimplePrediction is on the CF namespace, but it should be on models.

Update logo

Replace pylint with ruff

Create domain objects from tuples of primitives

https://github.com/trustyai-python/module/blob/5014edebeed5bfd00ed006d2dfaa43fab66ccb42/tests/test_counterfactualexplainer.py#L37

Infer domain type from instance type.

Refactor metrics namespaces to reflect exp-core

Refactor the metrics namespaces to reflect changes proposed in ADR-0002 and implemented in trustyai-explainability/trustyai-explainability#166