Balancing Fairness and Accuracy in AI: A Causal, Multi-Objective Perspective

wooden tray beside pots — Photo by Roman Kraft on Unsplash

Artificial Intelligence systems are no longer confined to research labs. They influence decisions about loans, employment, healthcare, and criminal justice—domains where fairness is not optional. Yet, much of modern machine learning still treats fairness as a secondary concern: something to be fixed after a model has already learned its patterns.

One of the most persistent assumptions in this space is that fairness and accuracy are inherently at odds. Improve one, and the other must suffer. But is this trade-off inevitable—or is it simply a limitation of how we frame the problem?

In our recent work, A Multi-Objective Approach to Balance Fairness and Accuracy, we argue for a different perspective: fairness should be treated not as a constraint or post-processing correction, but as a first-class optimisation objective, explored alongside accuracy rather than subordinated to it. I would like to congratulate my co-authors about this who are: 1. the doctoral candidate Zahid Irfan, Dr. Roisin Loughran, and Professor Fergal Mc Caffery. Basically this is the work was done by Zahid, who is a colleague as well as a very close friend of mine.

Why Bias Persists in Machine Learning

Bias in AI systems often reflects deeper structural issues: biased data collection, historical inequalities, and spurious correlations that models eagerly exploit. When these correlations involve protected attributes—such as sex, age, or race—the resulting systems may achieve impressive accuracy while still producing unfair outcomes.

Traditional bias-mitigation approaches typically fall into three categories:

Pre-processing, where the data is modified before training
In-processing, where fairness is incorporated into the learning algorithm
Post-processing, where predictions are adjusted after training

While all three have their place, many approaches operate largely as black-box fixes. They may improve a fairness metric, but often at the cost of interpretability and deeper understanding.

This is where causal modelling becomes essential.

Bringing Causality into the Picture

Correlation alone cannot tell us why a model behaves unfairly. Causal models, on the other hand, explicitly represent cause–effect relationships between variables.

We use Causal Bayesian Networks (CBNs)—directed acyclic graphs where nodes represent variables and edges encode causal influence. These structures allow us to reason about dependencies, confounders, and interventions, rather than relying solely on statistical association.

However, learning causal structures from data is a notoriously difficult problem. The search space of possible graphs grows exponentially, making exhaustive search infeasible.

To address this, we turned to Evolutionary Computation.

Evolving Causal Graphs with Grammatical Evolution

Our approach uses Grammatical Evolution (GE) to automatically generate and evolve causal graph structures. A context-free grammar constrains the search space to valid causal graphs, while still allowing a rich variety of structures to emerge.

Each individual in the evolutionary population represents a candidate causal graph. From this graph, we build a CBN, train it on data, and evaluate its performance.

Crucially, we do not evaluate performance using a single objective.

Fairness and Accuracy as Joint Objectives

Instead of collapsing everything into one score, we adopt a multi-objective optimisation framework using NSGA-II, a well-established evolutionary algorithm.

We optimise two objectives simultaneously:

Accuracy, measuring predictive performance
Fairness, measured using Equal Opportunity Difference (EOD), which captures disparities in true positive rates between protected groups

This produces not a single “best” model, but a Pareto front—a set of non-dominated solutions representing different fairness–accuracy trade-offs.

This is a powerful shift in mindset. Rather than asking “What is the best model?”, we ask:
“Which trade-off best fits the ethical and operational requirements of this domain?”

What We Observed

Using the German Credit dataset as a case study, our experiments showed that:

It is possible to achieve very low fairness disparity while maintaining competitive accuracy
Multiple causal graphs can yield similar performance, offering flexibility and interpretability
The evolved graphs are non-trivial, capturing meaningful dependencies among features
Practitioners can choose models that slightly sacrifice accuracy for substantial gains in fairness—or vice versa

Importantly, the causal graphs themselves provide insight. They allow us to inspect how features influence outcomes, opening the door to causal reasoning, domain validation, and future intervention analysis.