Advancing Equal Opportunity Fairness and Group Robustness through Group-Level Cost-Sensitive Deep Learning

Machine learning systems deployed in high-stakes decision-making scenarios increasingly face challenges related to fairness, spurious correlations, and group robustness.These systems can perpetuate or amplify societal biases, particularly affecting protected groups defined by sensitive attributes such as race or age.This paper introduces a novel cost-sensitive deep learning approach at different group levels that simultaneously addresses these interconnected challenges.Thus, our research uncovers a fundamental synergy between group robustness and group fairness.By developing a technique that enhances group fairness, we also improve the model's group robustness to spurious correlations.This approach encourages the model to focus on causally relevant features rather than misleading associations.We propose a comprehensive methodology that specifically targets group-level class imbalances, a crucial yet often overlooked aspect of model bias.By incorporating different misclassification costs at the group level, our approach, Group-Level Cost-Sensitive Learning (GLCS), provides a principled framework for handling both dataset-wide and group-specific class imbalances using different constraints in an optimization framework.Through targeted interventions for underrepresented subgroups, we demonstrate simultaneous improvements in equal opportunity fairness and worst-group performance, ensuring similar true positive rates across demographic groups while strengthening overall group robustness.Extensive empirical evaluation across diverse datasets (CelebA, UTKFace, and CivilComments-WILDS) demonstrates that our method effectively mitigates performance disparities and promotes more equitable outcomes without sacrificing overall model accuracy.These findings present evidence that addressing fundamental data distribution issues at the group level can naturally lead to fairer and more robust machine learning systems.Our work has significant implications for the ethical deployment of machine learning in critical domains such as healthcare, finance, and criminal justice, offering a practical path toward more equitable and reliable automated decision-making systems.