Direct Metric Optimization for Imbalanced Classification

Yash Travadi, Le Peng, Ying Cui, Ju Sun

Abstract: Imbalanced classification (IC) involves training models in datasets with uneven label distribution. It has been widely stuided [1] , [2] , with various applications such as object detection [3] , [4] , image retrieval [5] , [6] , and image segmentation [7] , [8] , [9] . In clinical research, data imbalance is also a prominent issue, where negative control usually significantly outweighs positive cases due to the innate low prevalence rate of certain diseases. Especially for tasks related to rare diseases (e.g., amyotrophic lateral sclerosis, where the prevalence rate can be as low as 0.001% [10] ), the imbalance problem is severe and almost inevitable.