Enhanced generalized normal distribution optimizer with Gaussian distribution repair method and cauchy reverse...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• High-dimensional datasets often contain noisy, redundant, and irrelevant features, degrading classification model performance.
• Feature selection is crucial for identifying optimal subsets, improving model efficiency and accuracy.
• Existing metaheuristic algorithms like Generalized Normal Distribution Optimization (GNDO) face challenges such as premature convergence and search imbalance.

Purpose of the Study:

• To propose a novel Binary Adaptive GNDO (BAGNDO) framework for effective feature selection.
• To enhance the efficacy of metaheuristic algorithms in addressing limitations of noisy, high-dimensional data.
• To improve classification accuracy and reduce feature subset size in machine learning models.

Main Methods:

• Developed the Binary Adaptive GNDO (BAGNDO) framework incorporating Adaptive Cauchy Reverse Learning (ACRL), an Elite Pool Strategy, and Gaussian Distribution-based Worst-solution Repair (GDWR).
• Evaluated BAGNDO's performance against nine state-of-the-art metaheuristic algorithms.
• Tested the framework on 18 UCI benchmark datasets using wrapper-based feature selection.

Main Results:

• BAGNDO achieved the highest classification accuracy on 14 out of 18 benchmark datasets.
• The framework consistently produced the most compact feature subsets compared to other algorithms.
• Statistical analyses (Wilcoxon signed-rank, Friedman tests) confirmed BAGNDO's significantly superior performance.

Conclusions:

• BAGNDO is a robust and efficient solution for wrapper-based feature selection in high-dimensional datasets.
• The proposed enhancements effectively balance exploration and exploitation, overcoming limitations of the original GNDO algorithm.
• BAGNDO offers a significant advancement in optimizing feature selection for classification tasks.