This repository contains the experimental results of our study on the impact of feature selection (FS) and feature extraction (FE) techniques in machine learning-based intrusion detection systems (IDS).
The experiments are conducted on three widely used benchmark datasets:
- UNSW-NB15
- AWID
- CSE-CIC-IDS2018
The study evaluates a comprehensive pipeline:
Balancing → Scaling → Feature Selection → Feature Extraction → Classification
The total pipeline configurations per dataset were explored, combining:
- Feature Selection (FS): Variance Threshold (VT), ANOVA F-test, Chi-Squared (Chi2)
- Feature Extraction (FE): PCA, LDA, ICA, SVD
- Classifiers (12 models): DT, LR, KNN, NB, SVM, RF, ET, GB, XGB, LGBM, CAT, STACK
The repository includes the following result files:
-
cv_results_summary.csv- Aggregated performance across all configurations
- Metrics reported as mean ± standard deviation
- Includes Macro-F1 and Weighted-F1 scores
-
cv_fold_level_records.csv- Detailed performance for each fold
- Enables reproducibility and statistical analysis
-
ablation_modellevel_records.csv- Model-level performance across FS–FE–Classifier combinations
- Includes Δ Macro-F1 to quantify performance changes
- LDA consistently provides the strongest performance across all datasets
- Stacking ensemble (STACK) achieves the best overall classification results
- FS→FE interactions are synergistic, not merely additive
- PCA may degrade performance, especially on smaller datasets (e.g., AWID)
- Larger datasets (UNSW, CIC) benefit significantly from supervised feature extraction (LDA)
The repository currently provides complete experimental results to support transparency and reproducibility of the reported findings.
👉 The source code and full implementation details will be released after the acceptance of the manuscript.