Unsupervised Feature Selection With Constrained ℓ₂,₀-Norm and Optimized Graph.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Traditional feature selection often involves separate steps for feature selection and similarity matrix construction.
• Existing sparse learning methods typically solve relaxed or approximate convex problems, potentially sacrificing accuracy.
• Unsupervised feature selection is crucial for dimensionality reduction and improving model performance without labeled data.

Purpose of the Study:

• To propose a novel unsupervised feature selection approach, RSOGFS, that unifies feature selection and similarity matrix construction.
• To directly address the l2,0-norm constrained problem for group feature selection, avoiding relaxation or approximation.
• To adaptively determine a similarity matrix that preserves the local manifold structure of data.

Main Methods:

• Developed a unified framework integrating feature selection and similarity matrix construction.
• Proposed a novel approach using constrained l2,0-norm (row-sparsity constrained) and optimized graph (RSOGFS).
• Introduced two optimization strategies to solve the original sparse constrained problem, providing convergence and approximation guarantees.

Main Results:

• The proposed RSOGFS method demonstrates superior performance compared to state-of-the-art methods on real-world datasets.
• Experimental results validate the effectiveness of the approach in solving a non-convex problem.
• The method adaptively constructs a similarity matrix that effectively preserves local data manifold structures.

Conclusions:

• The unified framework of RSOGFS offers a more effective approach to unsupervised feature selection.
• Directly tackling the l2,0-norm problem leads to improved performance over methods solving relaxed versions.
• RSOGFS provides a robust and theoretically sound method for feature selection and data representation.