Analysis of the Effect of Feature Selection on the Performance and Efficiency of XGBoost and Random Forest Algorithms in Phishing Site Detection

Purpose - This study examined how Recursive Feature Elimination (RFE) affects predictive performance, feature-selection stability, and computational efficiency in XGBoost and Random Forest phishing website classifiers. Design/methods/approach – A quantitative experiment analyzed 11,054 website observations with 30 predictors. Five feature configurations (30, 25, 20, 15, and 10) were compared using stratified five-fold cross-validation on an 80% development split, with RFE conducted inside the training portion of each fold. The configuration with the highest mean F1-score for each classifier was subsequently tested on an untouched 20% hold-out set. Findings – Random Forest with 25 features achieved 96.947% mean cross-validation accuracy and a 96.516% F1-score; XGBoost performed best with all 30 features (96.291% and 95.783%). Hold-out accuracies were 97.332% and 96.834%, respectively. The selected Random Forest configuration with 25 features required 10.402 seconds for model development, compared with 0.454 seconds for XGBoost using all 30 features. This descriptive comparison reflects differences in both model architecture and feature-selection treatment rather than the isolated computational effect of RFE. The small Random Forest improvement over its 30-feature baseline was descriptive and not tested for statistical significance. Research implications/limitations – RFE reduced training time after feature selection in some scenarios but increased total construction time. Inferences are confined to one benchmark dataset, fixed model settings, and one development split. Originality/value – This study provides an integrated empirical assessment of how RFE-based feature reduction affects predictive performance, feature-selection consistency, and computational costs in XGBoost and Random Forest. Its contribution lies in demonstrating that modest classification improvements may coincide with substantially higher model-development costs when feature-selection overhead is included.