Summary: In the field of fair supervised learning, ensuring that models do not rely excessively on sensitive attributes such as gender or ethnicity is a key challenge. One of the most widely used approaches to evaluate fairness in models is Demographic Parity (DP), which measures the independence between predictions and sensitive attributes. However, recent research has shown that using DP in learning algorithms can introduce inadvertent biases, especially when training data is imbalanced. This work analyzes the impact of DP-based regularization methods on the conditional distribution of predictions with respect to sensitive attributes. The results indicate that models may lean toward the majority category within the sensitive attribute, affecting system fairness. To mitigate this problem, a method based on Sensitive Attribute Distributionally Robust Optimization (SA-DRO) is introduced, which improves model robustness to variations in the distribution of the sensitive attribute. Through numerical experiments, the effectiveness of this methodology is demonstrated in both centralized and federated learning environments.
Introduction: In the development of Artificial Intelligence for critical decision-making, it is essential to ensure that algorithms are fair and do not generate discrimination based on sensitive attributes. Otherwise, there is a risk that models make biased decisions. This problem has sparked interest in the scientific community, which has proposed various strategies to improve fairness in machine learning models. In this context, Q2BSTUDIO, a company specialized in development and technological services, focuses on integrating advanced machine learning solutions with an ethical and responsible approach. Through the implementation of methodologies such as SA-DRO, our team seeks to ensure that models generate fair decisions, regardless of the sensitive attributes present in the data.
Related Work: Within the field of fair machine learning, various metrics have been explored to evaluate fairness in models, including mutual information, Pearson correlation, and kernel density estimation. At the methodological level, strategies to mitigate bias include preprocessing, postprocessing, and in-processing approaches. While preprocessing methods aim to transform data before training, postprocessing methods adjust decisions after inference. Our work focuses on in-processing approaches, which incorporate regularization during training to force the model to learn equitable patterns. Unlike other robust optimization approaches, SA-DRO uses the distribution of the sensitive attribute to control bias during training, achieving fairer models.
Conclusion: The analysis carried out demonstrates that DP-based supervised learning can induce unwanted biases when there is an imbalance in the training data. The proposed SA-DRO methodology mitigates these negative effects by optimizing system robustness against changes in the distribution of the sensitive attribute. Experimental results in centralized and federated learning environments validate the effectiveness of the approach. At Q2BSTUDIO, we are committed to developing fair technological solutions, ensuring that our algorithms are equitable in their performance. By addressing these challenges, we seek to contribute to a more transparent, responsible, and accessible AI ecosystem for all.
