Balancing Fairness and Accuracy Using Grammatical Evolution

Balancing predictive accuracy and fairness in machine learning is a critical challenge. In this work we explore Grammatical Evolution (GE) for learning causal graphs and applying interventions to improve interpretability, while maintaining fairness of the machine learning models. We focus on the Equalized Odds Difference (EOD) metric to evaluate fairness across different datasets, including the German Credit, COMPAS, and Adult datasets. The results demonstrate that GE can eectively learn causal structures that enhance fairness while maintaining competitive accuracy compared to baseline methods. Additionally, we introduce fitness functions GE-α, GE-β, and GE-γ. GE-α uses constraints to ensure fairness and accuracy are at an optimum level. GE-β employs a Minimax strategy that jointly minimizes unfairness and1-accuracy, ensuring balanced performance by reducing the worst-case metric, and GE-γ uses Average Causal Effect (ACE) based node dropping to enhance interpretability without compromising fairness and accuracy. Results demonstrate that the choice of fitness function significantly shapes the fairness-accuracy trade-off and GE can produce models with improved fairness while maintaining competitive accuracy.