Worst Group Performance

Worst-group performance, focusing on minimizing the error rate for the least-well-represented subgroup within a dataset, is a critical concern in machine learning. Research actively explores methods to improve this performance, often employing techniques like group distributionally robust optimization or two-stage approaches that leverage partially available group labels to guide model training. These efforts are driven by the need to mitigate biases and ensure fairness in applications where model predictions impact diverse populations, particularly in sensitive domains like healthcare. Current work investigates the influence of model size and architecture on worst-group generalization, aiming to develop more robust and equitable machine learning systems.