Unsupervised Feature Selection via Nonnegative Spectral Analysis and Redundancy Control.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• High-dimensional features in image processing are often redundant and noisy, hindering pattern recognition.
• Existing methods may struggle with noise and redundancy in feature representation.

Purpose of the Study:

• To propose a novel unsupervised feature selection scheme for images.
• To identify a discriminative subset of useful and redundancy-constrained features.
• To improve accuracy in image processing and pattern recognition tasks.

Main Methods:

• Developed a nonnegative spectral analysis with constrained redundancy (NSACR) method.
• Jointly leveraged nonnegative spectral clustering and redundancy analysis for feature selection.
• Employed row-wise sparse models with a general ℓ(2, p)-norm for noise robustness.

Main Results:

• The method simultaneously learns cluster labels and performs feature selection.
• Identified discriminative features while explicitly controlling redundancy.
• Achieved encouraging results on nine diverse image benchmarks (face, digit, object data).

Conclusions:

• The proposed NSACR algorithm is effective for unsupervised feature selection in images.
• Demonstrated superior performance compared to several representative algorithms.
• Highlights the importance of joint learning for discriminative feature selection.