Confidence-Driven Semi-Supervised Partial Label Learning
Chengkun Liu, Jun Zhang, Jing Chai · 2024
Semi-supervised Partial Label Learning (SPLL) aims to learn from a dataset comprised of both partial label examples each of which is associated with a candidate label set and unlabeled examples. The mainstream of SPLL methods usually construct a confidence matrix for training examples and by which to operate label disambiguation and classifier training. However, they treat examples with different confidence levels in the same strategy during training, which might result in degenerated learning performance. In this work, a novel method named COnfidence-DRiven (CODR) is proposed to deal with the above drawback. In specific, we iteratively update the confidence matrix and predictive network, and employ different strategies to deal with high-confidence and remaining low-confidence examples. Extensive experiments on real-world datasets demonstrate the superiority of CODR in classification accuracy compared with several other state-of-the-art methods.