International Journal of Innovative Information Systems

& Technology Research 14(4):227-242, Oct-Dec, 2026

 

  © SEAHI PUBLICATIONS, 2026    www.seahipublications.org    ISSN: 2467-8562

 

 

 

 

DOI:10.5281/zenodo.23951294

Design Of A Hybrid Filter-Wrapper Feature Selection Model Based On XGBoost For Next-Generation Network Security

 

 

Anas Sani,  Gilbert. I. O. Aimufua, Rashidah Funke Olanrewaju,  A. S. Abdullahi, Binyamin A. Ajayi

 

ABSTRACT

Modern networks, from 5G systems to the growing population of connected devices, generate far more traffic than security teams can inspect by hand, and the attacks hidden in that traffic are increasingly difficult to spot. Machine learning can flag malicious activity automatically, but its accuracy depends heavily on the features it learns from, and network records often carry dozens or hundreds of redundant or irrelevant features that slow the model down and cloud its judgement. This study addresses that problem with a hybrid filter-wrapper feature selection model built around XGBoost, in which a filter stage keeps only the features endorsed by both XGBoost feature importance and mutual information, a wrapper stage applies Recursive Feature Elimination with Cross-Validation (RFECV) to find the smallest subset that still gives the best accuracy, and the final XGBoost classifier handles class imbalance through scale_pos_weight rather than synthetic oversampling, with a 108-combination GridSearchCV routine provided for tuning on real-world data. Evaluated on four benchmark datasets spanning classic, modern, enterprise, and IoT environments (NSL-KDD, UNSW-NB15, CICIDS2017, and CICIoT 2023), the model achieved an average accuracy of 94.91 per cent, an average F1-score of 94.62 per cent, and an average AUC-ROC of 98.58 per cent, while cutting the number of features by between 53.5 and 79.2 per cent; on NSL-KDD it outperformed all six baseline classifiers, with McNemar's chi-squared test confirming statistically significant gains against five of the six. These results show that pairing a dual-criterion filter with a cross-validated wrapper yields an intrusion detection model that is both more accurate and more efficient than single-strategy approaches, and that generalises well across very different kinds of network traffic.

Keywords: intrusion detection, feature selection, filter-wrapper, XGBoost, mutual information, RFECV, network security, machine learning.