Procedural Fairness via Group Counterfactual Explanation

Researchers Gideon Popoola and John Sheppard have introduced a novel AI fairness framework called Group Counterfactual Integrated Gradients (GCIG), detailed in their arXiv paper "Procedural Fairness via Group Counterfactual Explanation." The work addresses a critical gap in machine learning fairness research, which has traditionally focused on outcome-oriented criteria like Equalized Odds while neglecting procedural fairness—how a model arrives at its predictions. GCIG is an in-processing regularization method that enforces explanation invariance across different protected groups, conditioned on the true label. For each input, it computes explanations relative to multiple Group Conditional baselines and penalizes cross-group variation in these attributions during the training process.

This approach formalizes procedural fairness as Group Counterfactual explanation stability, complementing existing fairness objectives that constrain predictions alone. The researchers empirically compared GCIG against six state-of-the-art fairness methods, and their results demonstrate that GCIG substantially reduces cross-group explanation disparity while maintaining competitive predictive performance and favorable accuracy-fairness trade-offs. The framework represents a significant advancement because it moves beyond simply ensuring equal outcomes to ensuring that the reasoning process itself is consistent and fair across different demographic groups.

The introduction of GCIG offers a principled and practical avenue for advancing fairness in AI systems beyond mere outcome parity. By aligning model reasoning across groups, the method helps build trust in AI systems, addressing concerns that different explanations for different protected groups can erode user confidence. The work has been submitted to ECML 2026 and represents an important step toward more transparent and equitable machine learning models.