Natural Visibility Graph (NVG) characterizes network traffic with topological descriptors that reflect structural properties. This work evaluates twenty‑one NVG‑derived metrics and investigates whether a compact subset can retain classification capability while reducing computational cost. Four importance analysis methods—SHAP, grouped permutation importance, Boruta and recursive feature elimination—are applied and their results merged by a consensus ranking strategy. According to the ranking, six configurations (Full21, Top15, Top10, Top7, Top5 and Top3) are built and tested on the CICIDS2018 dataset using a CNN classifier with stratified five‑fold cross validation. The three highest ranked metrics are avg_clustering_coeff_median, avg_clustering_coeff_std and avg_clustering_coeff_mean. Top3 achieves mean accuracy of 97.148 %, weighted F1 of 97.055 % and MCC of 0.9675, surpassing Full21’s 95.999 %, 95.521 % and 0.9549 respectively. The total runtime drops from 14961.39 s to 589.22 s, a reduction of 96.06 %. These findings indicate that importance‑guided metric reduction can provide a compact NVG representation with higher observed predictive performance and substantially lower computational cost under the evaluated setting.
Review