Intelligent algorithms for feature selection in supervised and unsupervised classification

Abstract

The growing availability of high-dimensional datasets across various domains has made feature selection a critical step in machine learning pipelines, as reducing irrelevant or redundant features enhances model interpretability and generalization. Due to the combinatorial nature of feature selection, traditional meth ods often lack the scalability and adaptability required for real-world problems. In response, this thesis investigates the application of intelligent metaheuristic algorithms to feature selection in both supervised and unsupervised learning settings. First, a comparative analysis of the Equilibrium Optimizer (EO) and Henry Gas Solubility Optimization (HGSO) algorithms is conducted for supervised classification tasks. Both algorithms are adapted to a binary feature space and evaluated on benchmark datasets using classification accuracy and feature reduction as performance criteria, highlighting their respective strengths and motivat ing a hybrid approach. Consequently, this thesis proposes HGSOEO, a hybrid algorithm that integrates the complementary exploration and exploitation capabilities of HGSO and EO. The proposed HGSOEO algo rithm is evaluated on the Twitter Spam Detection dataset and demonstrates superior performance in terms of classification accuracy and the number of selected features when compared to conventional metaheuristic and classical feature selection methods. Furthermore, the application of EO is extended to feature selection for clustering tasks, where labeled data are unavailable, by employing clustering validity criteria such as the Adjusted Rand Index (ARI) to guide the selection process. Experimental results across multiple datasets confirm the effectiveness and robustness of the proposed approaches. Overall, the findings of this thesis demonstrate that intelligent metaheuristic algorithms provide efficient and scalable solutions to the feature selection problem in both supervised and unsupervised learning contexts.