Training Batch — 8 Labeled Incidents
Shows the diversity needed in a real GNN-RCA training set.
Source File
Incident Summary
| # | ID | Fault Type | Root Cause | Alarms | UEs | Domain |
|---|---|---|---|---|---|---|
| 1 | INC00847291 | HW: SFP degradation | AGG router | 47 | 5,210 | Transport→RAN→Core |
| 2 | INC00852104 | Config: OSPF cost error | AGG router | 23 | 3,400 | Transport→RAN |
| 3 | INC00855678 | HW: Power supply | gNodeB | 12 | 2,800 | RAN |
| 4 | INC00861234 | SW: VNF memory leak | UPF | 35 | 15,000 | Core→RAN |
| 5 | INC00867890 | Congestion: Capacity | gNodeB | 18 | 45,000 | RAN→Core |
| 6 | INC00873456 | Config: Neighbor list | gNodeB | 8 | 800 | RAN |
| 7 | INC00879012 | Power: Commercial outage | CSG | 52 | 8,500 | Transport→RAN |
| 8 | INC00885678 | HW: Line card failure | Core router | 68 | 25,000 | Transport→RAN→Core |
Class Balance
- root_cause: 8 nodes (1 per incident) — ~8% of affected nodes
- Recommended handling: focal loss (γ=2), cost-sensitive weighting, oversampling via simulation
Training Split
Split by incident (not by node): 70% train / 15% validation / 15% test.