Direct pairwise AUC-surrogate training is attractive because it trains toward the ranking metric used at test time, especially under class imbalance. This creates a direct alignment between the training objective and the evaluation metric, which is a key advantage for ranking tasks.
However, the paper's title indicates that such training needs corrective power. The authors present a unified view of this requirement, suggesting that the direct alignment alone is insufficient and that additional correction is needed for effective optimization.
The abstract does not elaborate on the specific form of corrective power, but the unified view likely ties together existing approaches to AUC optimization under a common framework. This work highlights a subtle but important limitation in a popular training strategy.