Selective Classification Can Magnify Disparities Across Groups

A recent study by the Stanford Machine Learning Group reveals that selective classification, a technique allowing models to abstain from predictions when uncertain, can inadvertently worsen accuracy for specific data subpopulations. While this method generally improves average accuracy in high-stakes fields like medicine and natural language processing by deferring difficult cases to human experts, it fails to address disparities among minority groups. Using chest X-ray diagnosis for pleural effusion as a case study, researchers found that models relying on spurious correlations, such as the presence of chest drains, often misclassify untreated patients. Although overall performance metrics improve, accuracy for the clinically critical subgroup of untreated patients remains poor or deteriorates. Theoretical and empirical results suggest that selective classification is not an effective tool for resolving accuracy gaps between subgroups. Consequently, the authors warn practitioners about these failure modes and propose methods for developing more equitable selective classifiers to ensure fairer outcomes in real-world deployments.