Semisupervised Subspace Learning With Adaptive Pairwise Graph Embedding.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Graph-based semisupervised learning methods leverage data topology but face challenges with high-dimensional noisy features and Gaussian assumptions.
• Existing methods struggle to accurately represent data relationships and capture local submanifold structures, limiting representation discriminativeness.

Purpose of the Study:

• To propose a novel semisupervised subspace learning method, adaptive pairwise graph embedding (APGE), to address limitations in existing graph-based approaches.
• To improve the quality of constructed graphs and capture non-Gaussian local data structures for enhanced classification.

Main Methods:

• APGE constructs a k-nearest neighbor graph on labeled data to learn local discriminant embeddings, exploring non-Gaussian submanifold structures.
• A k-nearest neighbor graph is built on all samples and mapped to GE learning for adaptive global structure exploration.
• Adaptive neighborhood learning refines graph structure within an optimized subspace, ensuring robust learning of optimal graph and projection matrices.

Main Results:

• The method effectively explores local structure and enhances the discriminative ability of embedded data by clustering unlabeled data with labeled neighbors.
• A rank constraint on the Laplacian matrix clarifies graph structure and near-neighbor relationships, aligning connected components with the number of sample classes.
• Experiments on synthetic and real-world datasets demonstrate APGE's superior performance in exploring local structure and classification tasks.

Conclusions:

• APGE offers a robust and effective approach to semisupervised subspace learning by improving graph construction and capturing complex data structures.
• The method shows significant promise for improving classification accuracy and data representation in machine learning applications.