audit-AI: Ensuring Fairness in Machine Learning Models

audit

audit-AI: Ensuring Fairness in Machine Learning Models

A Python library for bias detection in AI models.

Access Platform

audit-AI: Ensuring Fairness in Machine Learning

In the rapidly evolving world of artificial intelligence, ensuring fairness and mitigating bias in machine learning models is crucial. audit-AI, developed by the Data Science team at pymetrics, is a Python library designed to address these challenges. Built on top of pandas and sklearn, audit-AI implements fairness-aware machine learning algorithms to detect and mitigate discriminatory patterns in training data and model predictions.

What is audit-AI?

audit-AI is an open-source tool that helps measure and mitigate the effects of bias in machine learning applications, particularly those involved in socially sensitive decision-making processes. It aims to provide a structured approach to understanding and improving the fairness of algorithms by identifying potential biases in training datasets and the resulting models.

Key Features

audit-AI offers a range of features to test for bias and audit algorithms:

  • Classification Tasks: Implements tests such as the 4/5th rule, Fisher's exact test, z-test, Bayes factor, and chi-squared test to evaluate bias in classification models.
  • Regression Tasks: Includes ANOVA, 4/5th rule, Fisher's exact test, z-test, Bayes factor, and chi-squared test for regression models.
  • Statistical and Practical Significance: Determines bias using statistical significance (p < .05) and practical significance (4/5ths rule).
  • Temporal and Regional Analysis: Uses the Cochran-Mantel-Hanzel test to check for differences over time or across regions.

Installation

To install audit-AI, you can use pip:

pip install audit-AI

Ensure you have the necessary dependencies installed, including scikit-learn, numpy, and pandas. For visualization, matplotlib and seaborn are recommended.

How to Use audit-AI

Here's a basic example of how to use audit-AI to test for bias:

from auditai.misc import bias_test_check
X = df.loc[:, features]
y_pred = clf.predict_proba(X)
# Test for bias
bias_test_check(labels=df['gender'], results=y_pred, category='Gender')

To visualize the results of different tests at various thresholds, use:

from auditai.viz import plot_threshold_tests
plot_threshold_tests(labels=df['gender'], results=y_pred, category='Gender')

Example Use Case

Consider a model predicting credit scores for a diverse population. audit-AI can help ensure that the model does not unfairly disadvantage any demographic group by comparing pass rates across different categories such as gender and ethnicity.

Compliance with Regulatory Standards

audit-AI aligns with the Uniform Guidelines on Employee Selection Procedures (UGESP), ensuring that all assessment tools treat protected groups fairly. It extends these principles to machine learning methods, providing a robust framework for bias detection and mitigation.

Conclusion

In a world increasingly reliant on automated decision-making, tools like audit-AI are essential for promoting fairness and accountability in AI systems. By integrating audit-AI into your machine learning workflow, you can take a proactive approach to identifying and addressing bias, ultimately leading to more equitable outcomes.

Explore audit-AI on GitHub and contribute to the ongoing effort to make AI fairer for everyone.