In high-stakes domains like medicine and law, the transition from research to deployment requires more than just high accuracy. As Umang Bhatt argues in Maintaining The Humanity of Our Models, models must be "humane," prioritizing the mitigation of human bias alongside predictive power. However, practitioners often encounter a frustrating technical tension: as they tighten fairness constraints to ensure equitable outcomes across demographic groups, the model's overall accuracy frequently drops.
This "accuracy-fairness trade-off" is the central hurdle in modern AI ethics. Recent research focuses on moving the Pareto frontier of this trade-off—finding ways to make models more equitable without sacrificing their utility.
The Source of the Tension
The conflict between accuracy and fairness usually arises because sensitive attributes (like race or gender) are often correlated with the target labels in historical data. If a model is trained to maximize accuracy on biased data, it will naturally "exploit" these correlations as predictive shortcuts.
For example, in healthcare, Shalmali Joshi and colleagues found in Machine learning is more accurate and biased than risk scoring tools... that while machine learning models outperform traditional risk scores in predicting surgical complications, they are significantly more biased. The model achieves higher precision by leaning into demographic patterns that reflect systemic inequities rather than purely biological risk.
Decoupling and De-correlation
To break this trade-off, researchers are developing methods to "decouple" sensitive attributes from the model's decision-making process.
One approach is causal de-correlation. In FITNESS: A Causal De-correlation Approach for Mitigating Bias in Machine Learning Software, Shangwen Wang and team use causal inference to identify and remove the links between sensitive features and target labels before training begins. Similarly, Bartosz Krawczyk and others introduced Fair OverSampling (FOS) in Towards a holistic view of bias in machine learning. FOS uses synthetic data generation to blur the features that allow a model to distinguish between protected groups, effectively forcing the model to learn more universal, unbiased features.
In-Process Mitigation and Game Theory
Rather than cleaning the data beforehand, "in-process" methods integrate fairness directly into the learning objective.
A notable example is the "Distraction" module proposed by Mehdi Yazdani-Jahromi in Distraction is All You Need for Fairness. This approach frames fairness as a game-theoretic "maximin" problem. By embedding specialized weights that act as a "distraction," the network is forced to ignore sensitive information while still focusing on the primary task.
In the realm of unsupervised learning, Haiming Xu and colleagues developed DFMVC-AKAN, which uses Kolmogorov-Arnold Networks and distribution alignment. The intuition here is to align the data distributions of different demographic groups in a latent space, ensuring that the clustering algorithm cannot "see" the sensitive attributes, thus preventing them from biasing the results.
Stability through Ensembles
Fairness metrics are notoriously "fickle"—they can fluctuate wildly based on how a dataset is split. Martin Hirzel and Samuel C. Hoffman explored this in An Empirical Study of Modular Bias Mitigators and Ensembles, demonstrating that while ensembles might not always increase the absolute level of fairness, they significantly reduce the volatility of fairness metrics.
This idea is extended to intersectional fairness—ensuring equity for groups defined by multiple attributes (e.g., Black women). Zhenpeng Chen and Jie Zhang argue in Diversity Drives Fairness that an ensemble of "mutant" models can provide more robust protection for these vulnerable subgroups than any single model could.
Leveraging the Unlabeled
Finally, the scarcity of labeled data for minority groups often exacerbates bias. Vahid Noroozi and Sara Bahaadini proposed SSFair in Leveraging Semi-Supervised Learning for Fairness using Neural Networks. By using pseudo-labeling on abundant unlabeled data, the model can learn a more representative distribution of the entire population, mitigating the impact of biased labels in small, non-representative training sets. This suggests that the path to fairness may not require sacrificing accuracy, but rather finding more creative ways to utilize all available information.