A neurodynamic optimization approach to supervised feature selection via fractional programming.

Area of Science:

• Machine Learning
• Data Mining
• Computational Neuroscience

Background:

• Greedy feature selection methods often lead to suboptimal results.
• Unsupervised redundancy minimization shows limited efficacy for classification tasks.
• Holistic feature selection requires balancing redundancy minimization and relevance maximization.

Purpose of the Study:

• To propose a neurodynamics-based holistic feature selection approach.
• To address the limitations of greedy and unsupervised methods in classification.
• To improve classification performance through optimized feature selection.

Main Methods:

• Defined an information-theoretic similarity coefficient matrix using multi-information and entropy.
• Formulated supervised feature selection as a fractional programming problem.
• Developed a neurodynamic approach using two one-layer recurrent neural networks.

Main Results:

• Demonstrated global convergence of the proposed neural networks.
• Showcased the superiority of the neurodynamic approach over existing methods.
• Achieved improvements in classification accuracy, precision, recall, and F-measure.

Conclusions:

• The neurodynamics-based holistic feature selection approach offers a robust solution.
• This method effectively minimizes feature redundancy while maximizing relevance for classification.
• The proposed approach provides a significant advancement in supervised feature selection.