The increasing sophistication and scale of malicious network activities demand a fundamental shift from traditional signature-based intrusion detection systems toward adaptive, data-driven security architectures. Machine learning (ML) provides an effective paradigm for addressing this challenge by identifying intricate and non-linear patterns associated with cyber threats within complex, high-dimensional network data. This study presents a comprehensive comparative analysis of four widely used ensemble and tree-based ML algorithms Random Forest (RF), Decision Tree (DT), XGBoost, and LightGBM applied to the multi-class classification of contemporary network intrusions. Using the benchmark CIC-IDS-2017 dataset, a meticulous preprocessing framework was implemented to ensure data integrity, reproducibility, and methodological rigor. Model performance was evaluated through standard classification metrics, with macro-averaged F1-score prioritized to provide an equitable assessment across highly imbalanced class distributions. Experimental findings reveal substantial differences in performance among the examined algorithms. Although RF, DT, and LightGBM achieved overall accuracy levels exceeding 99.8%, XGBoost consistently demonstrated superior capability in identifying minority attack categories, achieving the highest overall accuracy of 99.89% and a macro-averaged F1-score of 0.8903. These results highlight XGBoost’s enhanced generalization capacity and resilience to class imbalance, confirming its suitability for deployment in real-time cybersecurity environments. In conclusion, this research establishes a consistent methodological benchmark for evaluating ensemble-based intrusion detection algorithms. It underscores the critical importance of balanced model assessment in the context of skewed network traffic distributions. The findings suggest that XGBoost offers the most reliable and balanced performance profile for practical implementation within modern Security Operations Centers (SOCs), providing a strong foundation for adaptive and intelligent intrusion detection frameworks.
Özçelebi et al. (Tue,) studied this question.