Cross-Domain Applications of Machine Learning: A Comparative Case Study from Iris Classification to Infrastructure Assessment

Abstract

Machine learning has become foundational in data science for optimizing predictive and classification models across diverse domains. Ensemble methods—including Random Forest, gradient boosting algorithms (LightGBM, CatBoost), and stacking meta-models—demonstrate superior performance by capturing complex nonlinear relationships and managing high-dimensional feature spaces. However, comparative benchmarking across ensemble architectures remains limited, particularly regarding feature importance analysis and cross-domain methodology transferability. This study comprehensively evaluates four ensemble learning approaches— Random Forest, LightGBM, CatBoost, and a stacking meta-model with logistic regression meta-learner—on the Iris dataset (N=150), emphasizing classification performance and feature interpretability. Data preprocessing employed standardization and stratified 80/20 train-test splits, with 5-fold cross-validation ensuring generalization. Performance assessment utilized accuracy, precision, recall, F1-score, and ROC-AUC metrics. Results demonstrated Random Forest and LightGBM achieving 90.0% accuracy, while CatBoost and the stacking meta-model attained 93.3% accuracy. Feature importance analyses confirmed petal length (46.0%) and petal width (26.5%) as dominant discriminative features. Cross-domain comparison with infrastructure condition assessment revealed analogous ensemble superiority, validating methodology transferability across scientific domains. Future research should extend ensemble methodologies to high-dimensional datasets, integrating Neural Architecture frameworks for enhanced interpretability.