Discriminative and Geometry-Preserving Adaptive Graph Embedding for dimensionality reduction.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Dimensionality reduction is crucial for high-dimensional data analysis.
• Graph embedding methods aim to preserve data's geometric and discriminative information.
• Existing manual or automatic graph constructions alone have limitations in fully exploring data structure.

Purpose of the Study:

• To develop a novel framework, Discriminative and Geometry-Preserving Adaptive Graph Embedding (DGPAGE), for enhanced dimensionality reduction.
• To integrate manual and adaptive graph constructions into a unified framework.
• To improve the preservation of discriminative and geometric information in low-dimensional embeddings.

Main Methods:

• Developed DGPAGE, a unified graph embedding framework integrating manual and adaptive graph construction.
• Learned an adaptive graph by injecting information from predefined graphs.
• Jointly learned the adaptive graph with optimized projections for subspace generation.

Main Results:

• DGPAGE outperforms state-of-the-art graph-based dimensionality reduction methods on image datasets.
• Ablation studies confirm the benefits of integrating manual and adaptive graph construction.
• The proposed method generates embedded subspaces with superior pattern discrimination for image classification.

Conclusions:

• The integrated framework DGPAGE effectively preserves more discriminative and geometric information.
• DGPAGE achieves both adaptability and specificity, leading to improved performance.
• Combining manual and adaptive graph construction offers significant advantages in dimensionality reduction.