Neurodynamics-driven holistic approaches to semi-supervised feature selection.

Area of Science:

• Machine Learning
• Pattern Recognition
• Computational Neuroscience

Background:

• Feature selection is vital for machine learning, but supervised methods struggle with limited labeled data.
• Semi-supervised learning is challenging when labeled data is scarce compared to unlabeled data.

Purpose of the Study:

• To propose novel neurodynamics-driven holistic approaches for semi-supervised feature selection.
• To address the challenge of selecting relevant features with limited labeled data.

Main Methods:

• Defined information-theoretic semi-supervised similarity and relevancy measures using multi-information, entropy, and unsupervised symmetric uncertainty.
• Formulated fractional and iteratively weighted quadratic programming problems for feature selection.
• Developed three neurodynamic optimization approaches utilizing two projection neural networks.

Main Results:

• The proposed neurodynamic approaches demonstrated superior classification performance.
• Experimental results on six benchmark datasets validated the effectiveness against existing methods.

Conclusions:

• The neurodynamic-driven semi-supervised feature selection methods offer a robust solution for datasets with limited labels.
• These approaches effectively balance feature relevancy and redundancy for improved machine learning outcomes.