Unsupervised Feature Selection With Extended OLSDA via Embedding Nonnegative Manifold Structure.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Unsupervised feature selection often relies on spectral clustering to generate pseudo-labels.
• Existing methods face challenges with Laplacian matrix dependency and non-negative pseudo-labels.

Purpose of the Study:

• To propose a novel unsupervised feature selection approach addressing limitations of spectral clustering.
• To achieve non-negative pseudo-labels and enhance feature selection performance.

Main Methods:

• Extending orthogonal least square discriminant analysis (OLSDA) to the unsupervised domain.
• Imposing an orthogonal constraint on the class indicator to preserve manifold structure.
• Utilizing l2,1 regularization for row-sparse projection matrices, equivalent to l2,0 regularization.

Main Results:

• The proposed method successfully generates non-negative pseudo-labels.
• Experimental results on nine benchmark datasets demonstrate significant effectiveness.
• The approach overcomes the dependency on the Laplacian matrix inherent in spectral clustering.

Conclusions:

• The novel OLSDA-based unsupervised feature selection method is effective.
• It provides a robust alternative to spectral clustering for generating reliable pseudo-labels.
• The method enhances discriminative information extraction in unsupervised learning scenarios.