Class-Distribution-Aware Pseudo-Labeling for Semi-Supervised Multi-Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Pseudo-labeling is effective for leveraging unlabeled data in machine learning.
• Conventional methods struggle with semi-supervised multi-label learning (SSMLL) due to multiple labels and unknown label counts, leading to label inaccuracies.
• Existing approaches often introduce false positives or miss true positives in SSMLL.

Purpose of the Study:

• To propose a novel Class-distribution-Aware Pseudo-labeling (CAP) method for semi-supervised multi-label learning.
• To overcome limitations of conventional pseudo-labeling in handling multiple labels and unknown label counts.
• To improve the utilization of unlabeled data in complex multi-label classification tasks.

Main Methods:

• Introduced a regularized learning framework with class-aware thresholds for controlling pseudo-label assignment.
• Developed a class-distribution-aware thresholding (CAT) strategy to align pseudo-label distribution with true data distribution.
• Extended CAT into a label decision method for enhanced testing phase classification performance.

Main Results:

• Demonstrated that estimated class distribution is a reliable approximation even with limited labeled data.
• Showcased CAP's ability to effectively control positive and negative pseudo-label assignments per class.
• Empirically validated the efficacy of CAP on multiple benchmark datasets, confirming its superiority in SSMLL.

Conclusions:

• CAP effectively addresses the inherent challenges of semi-supervised multi-label learning.
• The proposed class-aware approach significantly improves pseudo-labeling accuracy and classification performance.
• Theoretical verification of estimated class distribution correctness and generalization error bounds support the method's robustness.