Benchmarking Counterfactual Explanation Methods for Network Resilience Classification Under Seismic Stress
IEEE ACCESS, cilt.14, ss.99586-99605, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3708897
- Dergi Adı: IEEE ACCESS
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.99586-99605
- Anahtar Kelimeler: Counterfactual explanations, Disaster management, Explainable AI (XAI), KPI, Network resilience, Telecommunication
- Gazi Üniversitesi Adresli: Evet
Özet
The cellular network continuity during earthquakes is critical for emergency coordination and public safety, yet predicting base station resilience at the individual cell level and deriving actionable recovery strategies remains largely unaddressed. This paper proposes a four-stage framework applied to operational KPI data from the Istanbul LTE network during the M6.1 earthquake of April 23, 2025, consisting of spatiotemporal feature engineering, hybrid resilience labeling, deep learning classification, and counterfactual explanation generation. The hybrid labeling strategy integrates 3GPP standard thresholds with a composite score aggregating 30 KPIs and reclassifies 26.44% of 50,413 analyzed cells relative to conventional assessment alone, uncovering 1,583 latently non-resilient cells whose capacity degradation is invisible to standard two-KPI criteria. Ten classifiers achieve balanced accuracies and F1 scores between 0.970 and 0.995. A systematic benchmark of six counterfactual methods across all ten classifiers identifies coverage as the operationally binding constraint. LORE is effectively disqualified, averaging 37.0% coverage with extreme instability (17.2%-86.0% across models), and PertCF achieves 100% coverage with the best proximity score of 24.5 and sparsity score of 34.1, producing sparse and field-deployable intervention targets that network engineers can act on without full reconfiguration. These findings offer a principled basis for post-disaster triage and resilience-focused infrastructure planning, with direct implications for 5G network management and disaster response policy.