Cognitively Enhanced Versions of Capuchin Search Algorithm for Feature Selection in Medical Diagnosis: a COVID-19...

Area of Science:

• Cognitive computation
• Machine learning
• Data mining

Background:

• Feature selection (FS) is vital in machine learning and data mining, requiring further research.
• Existing FS methods in cognitive computation have seen enhancements to improve performance.

Purpose of the Study:

• To introduce three adaptive versions of the capuchin search algorithm (CSA) with improved search capabilities.
• To utilize these adaptive CSA versions for optimal feature subset selection in conjunction with a k-Nearest Neighbor (k-NN) classifier.

Main Methods:

• Developed three adaptive CSA versions: exponential CSA (ECSA), power CSA (PCSA), and S-shaped CSA (SCSA).
• Incorporated strategies like automated control of inertia weight and acceleration coefficients to enhance CSA's search potency and convergence.
• Applied binary versions of the adapted CSAs with a k-NN classifier for feature selection on 24 benchmark datasets.

Main Results:

• The proposed adaptive CSA versions significantly outperformed the original CSA and other k-NN based FS methods.
• Binary ECSA achieved the best overall results, excelling in classification accuracy (18 datasets), specificity (13 datasets), sensitivity (10 datasets), and fitness values (14 datasets).
• Binary ECSA, PCSA, and SCSA demonstrated over 90% performance in specificity, sensitivity, and accuracy on multiple datasets.

Conclusions:

• The proposed adaptive CSA methods are efficient in improving classification accuracy by effectively exploring the feature space.
• These methods successfully identify the most relevant features for classification tasks, outperforming existing techniques.
• The adaptive ECSA shows particular promise for enhancing feature selection in machine learning applications.