Efficient Methods for Selective Classification under Balanced Error Rate

Deploying deep learning models in safety-critical tasks, like medical diagnosis, demands classifiers that can abstain from high-uncertainty samples to mitigate errors, which are known as selective classifiers. However, while these tasks often exhibit class imbalance, most existing approaches are based on conventional metrics such as accuracy, which are unsuited for imbalanced data. Recent work has proposed an algorithm that minimizes the balanced error rate, which is an appropriate metric for this case. Yet, their solution presents high complexity and suffers from poor scalability, restricting the application range due to computational costs. This work establishes sufficient conditions for an optimal selective classifier under the balanced error rate and proposes three novel methods that are fast and highly scalable. Experimental results show that the methods match or outperform the state-of-the-art algorithm on synthetic and real-world imbalanced datasets.