Unsupervised Feature Selection via Adaptive Graph Learning and Constraint.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Graph-based feature selection performance depends on similarity matrix quality.
• Existing methods often use fixed graphs, limiting adaptability to data structure.
• Inappropriate initial graphs can negatively impact the entire feature selection process.

Purpose of the Study:

• To propose a novel unsupervised feature selection method using adaptive graph learning and constraints (EGCFS).
• To select uncorrelated yet discriminative features by exploiting embedded graph learning and constraints.
• To address the limitations of fixed graph structures in traditional methods.

Main Methods:

• Developed an adaptive graph learning approach that incorporates similarity matrix structure into optimization.
• Obtained a closed-form solution for the graph coefficient through adaptive graph learning.
• Embedded a special graph constraint to probabilistically connect nearer data points.
• Integrated maximizing between-class scatter matrix with adaptive graph structure in a unified framework.

Main Results:

• The adaptive graph learning method learns graph structures dynamically.
• The embedded graph constraint enhances manifold structure and connects related data points.
• Experiments on benchmark datasets demonstrate the effectiveness and superiority of the EGCFS method.
• The method provides a unique perspective linking graph-based approaches and k-means clustering.

Conclusions:

• The proposed EGCFS method effectively addresses the limitations of fixed graphs in feature selection.
• Adaptive graph learning and embedded constraints lead to improved feature selection performance.
• The method offers a robust framework for unsupervised feature selection with enhanced structural understanding.