The Labeled Data Challenge
The single biggest barrier to production GNN-based RCA in telecom.
The Fundamental Problem
Network data is abundant — the problem is ground truth labels:
- ✅ Billions of PM counter samples, millions of alarms, complete topology
- ❌ “What was the ACTUAL root cause?” — this is what’s missing
Why Labels Are Hard
| Challenge | Impact |
|---|---|
| Root cause ≠ symptom | A fiber cut causes 500 RAN alarms, but the ticket says “RAN degradation” |
| Inconsistent ticketing | Different NOC engineers describe the same fault differently |
| Incomplete resolution | ~30% of tickets closed as “resolved” with no root cause explanation |
| Multi-cause incidents | Some outages have 2-3 contributing factors |
| Silent faults | Degradations that never trigger alarms or tickets |
| Cross-domain blindness | RAN/transport/core teams each see their piece — nobody labels end-to-end |
| Tribal knowledge | Senior engineers know patterns but never document them |
Practical Yield from Trouble Tickets
From 100,000 trouble tickets (typical):
- ~30,000 have extractable root cause info (30%)
- ~10,000 can be mapped to specific NEs + alarms (10%)
- ~5,000 are high-confidence clean labels (5%)
Class Imbalance
Telecom faults follow a heavy-tailed distribution. Root cause nodes are ~8% of affected nodes in any incident. The rarest faults (cascading failures, <1%) are often the most damaging.
Mitigations: Graph SMOTE, focal loss (γ=2), cost-sensitive weighting, over-sampling via simulation, hierarchical classification.
Related Concepts
- Labeling Strategies — 6 practical approaches
- GNN+DRL for RCA — the label-free alternative