Multi-Objective Fairness Approach Using Causal Bayesian Networks & Grammatical Evolution

Addressing unwanted biases has become critical as Artificial Intelligence systems are increasingly integrated into various aspects of society. Bias in decision-making can lead to unfair outcomes, perpetuating social inequalities and discrimination. Causal graphs enable the identification of causal mechanisms that may contribute to biased outcomes. Evolutionary computation techniques are well known for exploring large, complex solution spaces and evolving approximate optimal solutions over successive generations. We propose a novel approach that combines causal structures with grammatical evolution to create directed acyclic graphs for modelling and evolving solutions using fairness and accuracy as fitness criteria. Our approach evolves causal graphs that balance model fairness and performance in single-objective and multi-objective settings. Results show that the multi-objective optimization improved fairness by 32 percent while reducing accuracy by only 2.85 percent compared to the single-objective case. This demonstrates that integrating causal mechanisms with evolutionary computation can effectively develop Artificial Intelligence systems that are both accurate and fair.