Unsupervised feature selection algorithm based on L 2,p-norm feature reconstruction

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Traditional subspace feature selection methods use fixed distances, limiting adaptability and noise handling.
• Existing methods struggle with diverse datasets and are sensitive to outliers.

Purpose of the Study:

• Propose a novel unsupervised feature selection algorithm (NFRFS) for enhanced adaptability and performance.
• Address limitations of fixed-distance approaches in feature selection.

Main Methods:

• Introduced unsupervised feature selection algorithm based on [Formula: see text]-norm feature reconstruction (NFRFS).
• Employed a flexible p-norm for adaptable feature reconstruction and spatial distance representation.
• Integrated adaptive graph learning to preserve local data geometric structure.
• Utilized regularization constraints for sparse and low-redundancy feature selection.

Main Results:

• NFRFS demonstrated superior clustering performance across 14 benchmark datasets.
• Outperformed 10 existing unsupervised feature selection algorithms.
• The flexible norm approach enhanced adaptability to various data characteristics.

Conclusions:

• NFRFS offers an effective and adaptable unsupervised feature selection solution.
• Adaptive graph learning and flexible norms are crucial for robust feature selection.
• The proposed method shows significant promise for improving data clustering tasks.