xai-compare 0.1.1

This repository aims to provide tools for comparing different explainability methods, enhancing the interpretation of machine learning models.

---

Description

xai-compare is an open-source library that provides a suite of tools to systematically compare and evaluate the quality of explanations generated by different Explainable AI (XAI) methods. This package facilitates the development of new XAI methods and promotes transparent evaluations of such methods.

xai-compare includes a variety of XAI techniques like SHAP, LIME, and Permutation Feature Importance, and introduces advanced comparison techniques such as consistency measurement and feature selection analysis. It is designed to be flexible, easy to integrate, and ideal for enhancing model transparency and interpretability across various applications.

Installation

The package can be installed from PyPI:

Using pip:

pip install xai-compare

Explainers

xai-compare supports three popular model-agnostic XAI methods:

SHAP

• SHAP values provide global interpretations of a model's output by attributing each feature's contribution to the predicted outcome.
• Depending on the model type, the script initializes an appropriate explainer such as shap.TreeExplainer for tree-based models, shap.LinearExplainer for linear models, or shap.KernelExplainer for more general models. It then uses SHAP to analyze and explain the behavior of the model.

LIME

• LIME provides local interpretations of individual predictions by approximating the model's behavior around specific data points.
• The script initializes a LimeTabularExplainer and explains local predictions of the model using LIME.

Permutation Feature Importance

• Permutation Feature Importance assesses the impact of each feature on a model’s prediction by measuring the decrease in the model’s performance when the values of a feature are randomly shuffled.
• The script measures this dependency by calculating the decrease in model performance after permuting each feature, averaged over multiple permutations.

Comparison techniques

Feature selection

The FeatureSelection class in xai-compare is a robust tool for optimizing machine learning models by identifying and prioritizing the most influential features. This class leverages a variety of explainers, including SHAP, LIME, and Permutation Importance, to evaluate feature relevance systematically. It facilitates the iterative removal of less significant features, allowing users to understand the impact of each feature on model performance. This approach not only improves model efficiency but also enhances interpretability, making it easier to understand and justify model decisions.

Feature Selection Workflow

Consistency

The Consistency class assesses the stability and reliability of explanations provided by various explainers across different splits of data. This class is crucial for determining whether the insights provided by model explainers are consistent regardless of data variances.

Consistency Measurement Workflow

Sample notebooks

The notebooks below demonstrate different use cases for xai-compare package. For hands-on experience and to explore the notebooks in detail, visit the notebooks directory in the repository.

Feature Selection Test Notebook

Consistency Test Notebook

Main Demo Notebook

Call for Contributors

We're seeking individuals with expertise in machine learning, preferably explainable artificial intelligence (XAI), and proficiency in Python programming. If you have a background in these areas and are passionate about enhancing machine learning model transparency, we welcome your contributions. Join us in shaping the future of interpretable AI.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• The California housing dataset is sourced from scikit-learn.
• SHAP and LIME libraries are used for model interpretability.