Unsupervised feature selection with evolutionary sparsity.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• The ℓ_2,0-norm is crucial for unsupervised feature selection.
• Existing algorithms struggle with automatic sparsity determination and can converge to local optima, selecting less informative features.

Purpose of the Study:

• To propose a novel unsupervised feature selection method, Evolutionary Sparsity (EVSP).
• To address limitations of existing methods by automatically determining sparsity and avoiding local optima.

Main Methods:

• EVSP integrates feature selection with a sparse projection matrix and population search.
• A multi-objective evolutionary algorithm with binary encoding recursively determines optimal sparsity.
• A mutation-repair operator guides population evolution for high-quality solutions.

Main Results:

• EVSP effectively determines the optimal sparsity level.
• The method significantly outperforms several state-of-the-art unsupervised feature selection techniques.
• Experiments were conducted on eleven benchmark datasets with high dimensionality and sample size.

Conclusions:

• EVSP offers an effective approach to unsupervised feature selection by automatically optimizing sparsity.
• The proposed method enhances feature selection by avoiding trivial feature selection and improving overall performance.