Subspace Learning for Dual High-Order Graph Learning Based on Boolean Weight.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Subspace learning is crucial for unsupervised feature selection, identifying feature clusters to approximate original data spaces.
• Existing methods often overlook feature correlations and high-order neighborhood structures, limiting their ability to capture intrinsic data geometry.
• Graph-based approaches frequently focus on one-order neighborhoods, failing to preserve complex local geometric characteristics.

Purpose of the Study:

• To address limitations in current unsupervised feature selection methods.
• To propose a novel framework, subspace learning for dual high-order graph learning based on Boolean weight (DHBWSL).
• To enhance the exploitation of geometric structure information in dual spaces and preserve local geometric characteristics.

Main Methods:

• Developed a subspace learning framework incorporating dual-graph regularization to analyze geometric structures.
• Introduced dual high-order graphs with Boolean weights for adaptive selection of high-order adjacency matrices.
• Evaluated the proposed method on 12 public datasets against nine state-of-the-art algorithms.

Main Results:

• The proposed DHBWSL framework effectively integrates sample and feature relationships.
• Dual high-order graph learning with Boolean weights enhances the representation of original data spaces.
• Experimental results show DHBWSL significantly outperforms existing unsupervised feature selection algorithms.

Conclusions:

• DHBWSL offers a robust approach to unsupervised feature selection by leveraging dual high-order graph learning.
• The method successfully captures intrinsic spatial structures and preserves local geometric properties.
• DHBWSL demonstrates superior performance, providing a valuable advancement in the field.