Data Augmentation With Regularization for Multi-Labeled Complementary Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Multi-labeled complementary label learning (MLCLL) aims to reduce labeling costs in multi-label learning (MLL).
• Existing neural network models for MLCLL often overfit to noisy data, creating sharp decision boundaries.
• Label correlations are underexplored, limiting denoising capabilities in current MLCLL methods.

Purpose of the Study:

• To propose a novel framework, NMCB, to mitigate noisy information impact in MLCLL.
• To introduce the application of mixup techniques to the MLCLL problem for improved robustness.
• To enhance generalization ability and reduce sensitivity to noisy labels in MLCLL models.

Main Methods:

• A tailored mixup strategy is employed to create smoother decision boundaries.
• A model is developed to automatically extract label correlations from transformed non-complementary labels.
• Consistency regularization is utilized with extracted correlations as alignment objectives for instance augmentations.

Main Results:

• The proposed NMCB framework demonstrates reduced sensitivity to noisy labels.
• Smoother decision boundaries lead to enhanced generalization ability.
• Exploration of label correlations further improves model performance in MLCLL tasks.

Conclusions:

• NMCB effectively alleviates the impact of noisy information in MLCLL.
• The integration of mixup and label correlation extraction offers a promising direction for robust MLCLL.
• Empirical results validate the effectiveness of the NMCB framework.