Abstract The rapid increase in data volume across modern applications has resulted in datasets with very high dimensionality, creating challenges such as increased computational cost, overfitting, and limited interpretability. Feature selection addresses these issues by identifying a subset of informative attributes while preserving the original meaning of the data 1, 10. Among existing techniques, wrapper and embedded methods are widely used due to their strong integration with predictive models 2, 7. This paper presents a comparative analysis of wrapper and embedded feature selection approaches for high-dimensional datasets, focusing on their principles, computational complexity, scalability, and applicability across domains. The discussion draws on established theoretical foundations and empirical studies to provide guidance for selecting suitable feature selection strategies in complex learning environments.
Kamble et al. (Sat,) studied this question.