Semi-Supervised Learning Through Label Propagation on Geodesics.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Graph-based semi-supervised learning (SSL) is a significant area of research.
• Existing methods face challenges in constructing effective graphs for complex data distributions and utilizing pair-wise similarity for robust label propagation.

Purpose of the Study:

• To develop an effective graph construction method for data on multiple manifolds.
• To define and employ a novel pair-wise similarity measure for robust label propagation.
• To enhance classification performance by integrating data structure and label information.

Main Methods:

• A simple and effective graph construction method ensuring data point connectivity.
• Characterization of global pair-wise data similarity using geodesic distance approximated by graph distance.
• Utilizing Kullback-Leibler divergence to measure similarity inconsistency and introducing a novel regularization term for margin optimization.

Main Results:

• The proposed method demonstrates superior data similarity compared to Euclidean distance-based methods.
• The approach effectively utilizes data structure and label information for improved classification.
• Efficient optimization and convergence analysis are provided, along with out-of-sample extension capabilities.

Conclusions:

• The developed graph-based SSL method offers a robust solution for complex data distributions.
• The novel similarity measure and regularization term significantly enhance label propagation and classification accuracy.
• The method shows strong performance against state-of-the-art SSL techniques in image classification tasks.