Optimizing Weighted Random Forests for Phishing Website Detection

Phishing attacks remain a prevalent threat on the web, and reliable detection systems rely on learning discriminative patterns from annotated data. In this work we conduct a comparison of ten classification methods on the widely used UCI Phishing Websites dataset. Our goal is twofold: (i) to establish a performance baseline across diverse algorithms and metrics, and (ii) to evaluate a lightweight optimization of weighted random forests that combines mutual-information based feature selection with threshold tuning to maximize the F1-score. We perform stratified five-fold cross-validation repeated over three random seeds and report accuracy, precision, recall, F1, area under the ROC curve (AUC) and computational costs. The proposed Optimized Weighted Random Forest (OWRF) achieves a competitive F1 of 0.962±0.004, but a standard Random Forest baseline attains 0.975 ± 0.002, suggesting that feature selection and threshold optimisation are insufficient to outperform strong ensembles on this dataset. Paired Wilcoxon signed-rank tests reveal statistically significant differences between OWRF and all baselines.