Complementary label learning with scarce annotated information

Bing Zhao

doi:10.1016/j.knosys.2025.114887

Complementary label learning with scarce annotated information

2025-11-13

0

PRE

AI

DOI：10.1016/j.knosys.2025.114887

Original

Original request for help

Save

Abstract

En 中文

Collecting accurately labeled data is widely recognized as costly and constitutes a critical challenge in classification tasks. Therefore, complementary label learning (CLL), which assigns incorrect labels to samples, has attracted increasing attention for its ability to reduce annotation costs. However, in large-scale data scenarios, the assignment of complementary labels still faces issues such as low efficiency and class imbalance. To this end, this paper proposes a novel CLL framework (TCU-CLL) for scarce annotated information, aiming to achieve accurate classification by relying on limited true and complementary labels. Unlike existing CLL methods, TCU-CLL fully exploits the typically ignored low-confidence information to automatically generate complementary labels for abundant unlabeled samples, thus improving the labeling efficiency. Subsequently, to mitigate the potential complementary label class imbalance and overfitting problems, the Kullback-Leibler divergence balancing technique and consistency regularization strategy are employed, respectively. Ultimately, an estimation error bound is derived to theoretically guarantee model convergence. Extensive numerical experiments further validate the effectiveness of the proposed method. Specifically, under the annotation ratio of approximately 1 %, TCU-CLL consistently delivers competitive classification results on benchmark datasets.

Complementary label learning with scarce annotated information

Abstract

Journal

Organization