FGMPSO: a hybrid firefly-gradient-MOPSO framework for high-dimensional feature selection

Abstract

When working with high-dimensional datasets, selecting the most relevant features is essential for improving both model clarity and processing efficiency, all while keeping predictive accuracy intact. In response to this challenge, the study introduces firefly-gradient-multi-objective particle swarm optimization (FGMPSO), an advanced hybrid technique that blends the firefly algorithm, gradient descent (GD), and multi-objective particle swarm optimization (MOPSO). This approach is specifically designed to identify an optimal subset of features that balances dimensionality reduction with strong classification performance. The method was evaluated on eight benchmark datasets and compared against multiple PSO-based feature selection techniques. The empirical results demonstrated that FGMPSO consistently achieved superior or competitive classification accuracy while selecting significantly fewer features. Notably, in several datasets, FGMPSO not only reduced dimensionality but also outperformed other methods in terms of classification accuracy. This efficiency is attributed to the intelligent exploration of the search space by the firefly algorithm, refinement via GD, and effective trade-off optimization enabled by MOPSO. The findings suggest that FGMPSO is a robust and scalable solution for feature selection, particularly suitable for complex and high-dimensional datasets. Its adaptability, convergence speed, and balance between dimensionality reduction and accuracy position it as a valuable tool in modern machine learning pipelines.