.png){kind=icon}
.png){kind=icon}
Return
A multi-class partial hinge loss for partial label learning
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-023-04954-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As an important branch of weakly supervised learning, partial label learning (PLL) tackles the problem where each training instance is associated with a set of candidate labels, among only one is correct. Most existing PLL algorithms elaborately designed loss functions and update strategies to learn potential ground-truth labels among candidate labels with deep neural networks. However, these algorithms are susceptible to the cumulative error caused by noisy label propagation when updating label confidences, this will make the deep models tend to overfit the noisy labels, thereby achieving poor generation performance. To remedy this issue, we propose a general framework multi-class partial hinge loss (MPHL) for PLL, which can disambiguate the candidate labels by optimizing the margin between the maximum modeling output from partial labels and that from non-partial ones. More importantly, the partial hinge loss can adaptively optimize the separation hyperplane to reduce the influence of cumulative error. Meanwhile, we introduce graph laplacian regularization to full mine the relationship between candidate labels of similar instances to constrain the separation hyperplane to improve the robustness of disambiguation. Extensive experimental results demonstrate that the multi-class partial hinge loss significantly outperforms the state-of-the-art counterparts.
Keywords:
Partial label learning
Multi-class partial hinge loss
Graph laplacian regularization
Loss function
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W