Optimal Graph Learning-Based Label Propagation for Cross-Domain Image Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Label Propagation (LP) is a semi-supervised learning method relying on similarity graphs.
• Cross-domain problems, where training and test data distributions differ, degrade LP performance.
• Existing LP methods struggle with domain shift due to weak inter-domain connections.

Purpose of the Study:

• To propose Optimal Graph Learning-based Label Propagation (OGL2P) for robust cross-domain learning.
• To enhance label propagation by optimizing both cross-domain and intra-domain graph structures.
• To improve feature extraction for better domain invariance and local discriminability.

Main Methods:

• OGL2P optimizes three graphs: one cross-domain and two intra-domain.
• Label propagation leverages these graphs to connect similar samples across domains and preserve domain-specific structures.
• Graph embedding optimizes graphs in a subspace for noise robustness and feature extraction.

Main Results:

• OGL2P demonstrates improved insensitivity to cross-domain problems.
• The method extracts locally discriminative and domain-invariant features.
• Experiments show OGL2P outperforms state-of-the-art cross-domain approaches on five datasets.

Conclusions:

• OGL2P effectively addresses the challenges of cross-domain label propagation.
• The proposed graph optimization strategy enhances robustness and accuracy.
• OGL2P offers a promising solution for semi-supervised learning in heterogeneous data environments.