Supervised Feature Selection via Collaborative Neurodynamic Optimization.

Area of Science:

• Machine Learning
• Pattern Recognition
• Optimization

Background:

• Feature selection is vital for machine learning and pattern recognition, aiming to identify informative features.
• Existing methods often involve complex optimization formulations.

Purpose of the Study:

• To formulate supervised feature selection as a mixed-integer and biconvex optimization problem.
• To propose two novel collaborative neurodynamic optimization (CNO) approaches for solving these problems.

Main Methods:

• Formulating feature selection as a mixed-integer optimization problem with weighted redundancy and relevancy.
• Reformulating the problem into a bound-constrained biconvex optimization problem.
• Developing two CNO approaches using recurrent neural networks (RNNs) and projection networks.

Main Results:

• The proposed CNO approaches effectively solve the formulated feature selection problems.
• Experimental results on 13 benchmark datasets demonstrate superior performance.
• The methods achieved higher average classification accuracy compared to mainstream techniques.

Conclusions:

• The developed CNO approaches offer a superior method for supervised feature selection.
• These techniques enhance classification accuracy in machine learning applications.