PreCoF: counterfactual explanations for fairness.

Area of Science:

• Artificial Intelligence
• Machine Learning Ethics
• Algorithmic Fairness

Background:

• Machine learning models in high-stakes decisions risk amplifying dataset bias.
• Lack of a universal metric for detecting model bias hinders fairness assessments.
• Understanding bias nature is crucial for selecting appropriate mitigation strategies.

Purpose of the Study:

• Integrate Explainable AI (XAI) with fairness research to provide insights into model bias.
• Develop a novel metric, Predictive Counterfactual Fairness (PreCoF), to detect both explicit and implicit bias.
• Assess the effectiveness of counterfactual explanations in identifying unfair discrimination.

Main Methods:

• Utilizing (Predictive) Counterfactual Explanations to analyze model behavior.
• Developing and applying the PreCoF metric for bias detection.
• Comparing attribute presence in explanations for protected versus unprotected groups.

Main Results:

• The PreCoF metric successfully detects implicit bias by analyzing attribute importance in counterfactual explanations.
• Identified instances where models disadvantage protected groups through correlated attributes, not direct sensitive attribute use.
• Demonstrated the utility of counterfactual explanations in uncovering subtle forms of algorithmic discrimination.

Conclusions:

• PreCoF offers a robust method for assessing algorithmic fairness, particularly for implicit bias.
• Counterfactual explanations are valuable tools for auditing machine learning models for fairness.
• Findings can inform policymakers on the justification of discriminatory outcomes in automated decision-making.