Answers are grounded in this paper and cite their sources.
Network intrusion detection relies on feature selection to control dimensionality across severely imbalanced traffic classes, and metaheuristic wrappers such as the grey wolf optimiser dominate this literature.Seeding such a search with feature importance rather than at random is an attractive refinement, and informed initialisation is well supported elsewhere; that literature also warns that importance seeding which ignores subset size and inter-feature redundancy is its principal failure mode, and this has not been quantified for intrusion detection.We implement SHAPinitialised binary grey wolf optimisation (SHAP-BGWO), drawing each initial mask as xj ∼ Bernoulli(φj/maxk φk) from mean absolute SHAP values, and evaluate a 2 × 2 ablation separating the warm start from a SHAP-coherence fitness term against a no-selection control, five conventional selectors and the reimplemented selection stages of two recent methods, on CIC-IDS2017 and UNSW-NB15 over 10 seeds with an audited train/test firewall, BCa intervals and Holm correction.Maximum normalisation collapses the expected initial subset to 8.1 of 78 features on CIC-IDS2017 and 4.9 of 42 on UNSW-NB15.On CIC-IDS2017, SHAP seeding changes macro-F1 by -18.67 points without the coherence term and by -11.25 points with it; neither contrast is significant at ten seeds, and no arm reaches Boruta (0.8000) or the no-selection control (0.7767) in mean macro-F1.On UNSW-NB15 the two contrasts are -11.56 and -12.49points and both are significant after Holm correction.A corrected re-run that selects the ensemble's class leaders on a training-side calibration split, under a prior that collapses the initial subset far less (21.1 of 78 and 13.5 of 42 features), shows no comparable loss, which locates the failure in the collapse of the initial subset rather than in importance seeding as such.Importance-proportional seeding needs an explicit size or redundancy constraint to be viable.
No takes yet. Share an insight, caveat, or question.
Thannon et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: