Probabilistic Semi-Supervised Learning via Sparse Graph Structure Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Semi-supervised learning (SSL) often relies on predefined graphs or assumptions like locally linear embedding.
• Existing SSL methods struggle with high-dimensional, noisy data and limited labeled examples.

Purpose of the Study:

• To develop a probabilistic SSL framework capable of learning sparse graph structures from high-dimensional, noisy data.
• To enhance classification performance using minimal labeled data by integrating density estimation and distance preservation.

Main Methods:

• A probabilistic framework employing sparse graph structure learning.
• A unified model for density estimation and pairwise distance preservation, robust to data noise.
• Leveraging labeled data to guide density estimation and refine graph structure.

Main Results:

• The proposed SSL model successfully learns sparse weighted graphs from unlabeled and limited labeled high-dimensional data.
• Demonstrated robustness to input data noise through expectation-based distance calculations.
• Achieved promising results and significant improvements in SSL settings compared to existing methods, especially with small labeled datasets.

Conclusions:

• The developed framework offers a robust and effective approach to semi-supervised learning, particularly for high-dimensional and noisy datasets.
• The method shows strong potential for improving classification accuracy when labeled data is scarce.
• The integration of density estimation and graph learning provides a powerful tool for uncovering underlying data structures.