Features Selection: a Systematic Comparison of Traditional, Metaheuristic, Hybrid, and Deep Learning Approaches

Feature selection plays a vital role in optimizing machine learning workflows, especially when dealing with highdimensional datasets where redundant or irrelevant features degrade model performance and increase computational complexity. This study presents a comprehensive comparative analysis of feature selection methods applied to the imputed Wild Bird Highly Pathogenic Avian Influenza (HPAI) dataset, a critical resource for avian disease surveillance and outbreak prediction. We evaluated traditional, metaheuristic, hybrid and deep learningbased FS approaches under a unified experimental setup. The data set was pre-processed and imputed to address missingness patterns, followed by optimizing feature subsets using various FS strategies. Benchmark classifiers, namely Random Forest and Logistic Regression, were employed to assess the predictive utility of the selected feature subsets. The results demonstrate that hybrid and deep learning-based methods, particularly Transformerbased FS, significantly outperform conventional techniques in terms of F1-Score, dimensionality reduction, and stability, albeit with higher execution times. This paper highlights the strengths and limitations of each Feature selection family, providing insights into their applicability to epidemiological modeling and real-world surveillance datasets.