Robust AUC maximization for classification with pairwise confidence comparisons

Supervised learning has been wildly applied to various realistic applications, such as healthcare (e.g., disease diagnosis) and transportation (e.g., autonomous vehicles) Unfortunately, to train an effective model, it often requires a large number of labeled examples, which has become a critical bottleneck in the case that manual annotating the class labels is costly.

To solve the problem, a research team led by Shengjun HUANG published their new research 15 August 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team focused on a recently proposed weakly-supervised learning framework called pairwise comparison (Pcomp) classification. In Pcomp framework, instead of pointwise labeled example, the training data is given in the form of pairwise comparison data, which represents one of two examples is more likely to be positive.

The team proposed the Pcomp-AUC (PC-AUC for short) framework for robust AUC maximization by minimizing pairwise surrogate loss with only Pcomp data. Theoretically, they disclosed that the proposed PC-AUC metric is linearly proportional with respect to AUC, showing that maximizing PC-AUC is equivalent to maximizing AUC. Comprehensive experimental results on multiple datasets demonstrate the effectiveness of the proposed method.

For the future direction, one is to improve the performance of Pcomp classification by further leveraging the information of unlabeled data.

DOI: 10.1007/s11704-023-2709-5