Answers are grounded in this paper and cite their sources.
Imbalanced multiclass network intrusion detection presents three coupled challenges: dominant classes can absorb structurally similar rare attacks, global reweighting cannot target the resulting class-specific confusion, and unconstrained specialist models can distort evidence for unrelated attack families. We propose hierarchy-constrained expert boosting (HCE-Boost), a fully supervised ensemble that addresses these challenges through three coordinated modules. First, a leakage-controlled protocol fixes the training, validation, and test roles before any transformation or model selection. Second, equally weighted XGBoost and LightGBM experts estimate the global class distribution, while validation errors activate a local specialist only for a semantic family or rare-absorption group that satisfies fixed criteria. Third, a probability-mass-conserving output rule allows the specialist to revise within-group proportions without increasing the total probability assigned to that group. Under fixed, class-complete evaluation with five matched seeds, HCE-Boost attained mean macro-averaged F1 scores (macro-F1) of 64.99%, 93.73%, and 92.39% on NSL-KDD, the supplied CIC-IDS2017 subset, and edge-IIoTset, respectively. These values were the highest among eight methods evaluated under the same protocol and exceeded the strongest corresponding baselines by 2.20, 0.71, and 0.28 percentage points. Component analysis showed a consistent specialist benefit on NSL-KDD, an inconclusive mean benefit on CIC-IDS2017, and no specialist activation on Edge-IIoTset. The results support validation-governed hierarchical correction as a method for improving minority-sensitive classification within the evaluated fixed-split setting. They do not establish operational edge deployment, which requires model compression and hardware-level latency, memory, throughput, and power measurements.
No takes yet. Share an insight, caveat, or question.
Renjia Zhang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: