Nonlinear Locality-Preserving Projections With Dynamic Graph Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Locality-Preserving Projections (LPPs) rely on affinity graphs to represent local manifold structure.
• Existing LPP methods suffer from pre-designed graphs inconsistent with data distribution and linear projections damaging nonlinear structures.

Purpose of the Study:

• To propose a nonlinear dimensionality reduction model, Deep Locality-Preserving Projections (DLPPs), addressing limitations of current LPP approaches.
• To simultaneously improve affinity graph quality and preserve nonlinear manifold structures in low-dimensional representations.

Main Methods:

• DLPPs utilize deep autoencoders (AEs) within two distinct loss functions to extract discriminative features.
• The first loss function adaptively determines sample affinities in an intermediate layer via nonlinear mapping, preserving manifold structure.
• The second loss function learns affinities in the reconstruction layer, ensuring denoised samples maintain a good manifold structure.

Main Results:

• The proposed DLPP model effectively preserves nonlinear manifold structures in low-dimensional spaces.
• Adaptive affinity graph learning reduces sensitivity to initial weights and avoids noisy/redundant features.
• Experiments on toy and benchmark datasets validate the model's effectiveness in nonlinear dimensionality reduction.

Conclusions:

• DLPPs offer a robust solution for nonlinear dimensionality reduction by integrating adaptive graph learning and deep feature extraction.
• The model successfully minimizes manifold structure mismatches between denoising and low-dimensional spaces.
• DLPPs represent a significant advancement over traditional LPP methods for complex data analysis.