Computational Resource Optimisation in Feature Selection under Class Imbalance Conditions

Feature selection is crucial for reducing data dimensionality as well as enhancing model interpretability and performance in machine learning tasks. However, selecting the most informative features in large dataset often incurs high computational costs. This study explores the possibility of performing feature selection on a subset of data to reduce the computational burden. The study uses five real-life datasets with substantial sample sizes and severe class imbalance ratios between 0.09 – 0.18. The results illustrate the variability of feature importance with smaller sample fractions in different models. In this cases considered, light gradient-boosting machine exhibited the least variability, even with reduced sample fractions, while also incurring the least computational resource.