Kernel-Based Representation Alignment for Class Imbalanced Semi-Supervised Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semi-supervised learning (SSL) leverages unlabeled data to overcome limitations of scarce labeled data.
• Real-world scenarios often present domain shifts and imbalanced class distributions, challenging standard SSL methods.
• Ensuring consistent class representations is crucial for machine learning robustness with imbalanced datasets.

Purpose of the Study:

• To develop a novel kernel function mapping strategy for semi-supervised learning.
• To enhance machine learning model robustness against imbalanced datasets and domain shifts.
• To refine pseudo-label prediction at the representation level for improved accuracy.

Main Methods:

• Proposed a kernel function mapping strategy using a Gaussian kernel.
• Mapped unlabeled data representations to labeled data centroids in an infinite-dimensional space.
• Implemented a selective strategy to correct majority class predictions while preserving minority class confidence.

Main Results:

• The proposed method effectively aligns class representations, improving robustness to imbalanced data.
• Refined pseudo-labeling at the representation level led to superior performance.
• Demonstrated superior performance compared to state-of-the-art approaches across various benchmarks.

Conclusions:

• The kernel function mapping strategy offers a straightforward yet effective solution for SSL with imbalanced data.
• The approach enhances machine learning model reliability in real-world applications with skewed data distributions.
• Validated effectiveness through extensive evaluations on diverse benchmarks and training settings.