Katja Bengeri, Junoš Lukan and Mitja Lustrek

Abstract
Feature selection is a crucial step in building effective machine
learning models, as it directly impacts model accuracy and interpretability. Driven by the aim of improving stress prediction
models, this article evaluates multiple approaches for identifying the most relevant features. The study explores filter-based
methods that assess feature importance through correlation analysis, alongside wrapper methods that iteratively optimize feature
subsets. Additionally, techniques such as Boruta are analysed for
their effectiveness in identifying all important features, while
strategies for handling highly correlated variables are also considered. By conducting a comprehensive analysis of these approaches, we assess the role of feature selection in developing
stress prediction models.