Labeling Strategies for Telecom GNN RCA

Six approaches, from most common to most advanced.

Strategy 1: Trouble Ticket Mining (Most Common, Most Noisy)

Extract root cause labels from ServiceNow/Remedy tickets using NLP:

1. NLP entity extraction (equipment names, interfaces, fault types)
2. Cause-effect relation extraction
3. Structured RCA label output

Yield: ~5% clean labels from raw tickets. Challenges include vague descriptions (~40%), wrong categorization (~20%), copy-paste templates (~15%).

Strategy 2: Expert Annotation (Highest Quality, Lowest Scale)

SMEs manually review historical incidents with full context (alarm timeline, KPI dashboards, topology view).

Cost: $50K–$200K for 1,000 labels. 30-90 minutes per incident. Inter-annotator agreement: 70-85%.

Use for: bootstrapping initial model, gold-standard test sets.

Strategy 3: Fault Injection / Chaos Engineering

Deliberately inject known faults into lab/digital twin. Perfect ground truth, unlimited scale, but synthetic ≠ real.

Tools: Simu5G (OMNeT++), ns-3, Cisco CML/GNS3, LitmusChaos, Keysight/Spirent.

Strategy 4: Semi-Supervised & Self-Supervised Learning

Leverage vast unlabeled data with few labels:

• Masked node prediction (GraphMAE) — predict masked KPIs
• Link prediction — predict removed edges
• Graph contrastive learning — learn robust representations
• Label propagation — GNNs naturally propagate sparse labels through graph topology

GNNs are uniquely label-efficient because message passing propagates labels through the network structure.

Strategy 5: Active Learning

Let the GNN select the most uncertain predictions for human labeling. Reaches ~85% accuracy with only 15% of data labeled (vs. 40%+ for random sampling).

Strategy 6: Causal Discovery

Learn causal relationships from observational data using Granger causality, PC algorithm, or DAG-GNN. No labels needed — discovers cause-effect structure directly.

Combined Production Pipeline

1. Bootstrap (Month 1-3): Fault injection (2K labels) + expert annotation (500) + NLP mining (3K noisy)
2. Train (Month 3-4): Self-supervised pre-training + semi-supervised fine-tuning on ~5K labels
3. Refine (Month 4+): Active learning from NOC feedback (20 labels/week)
4. Closed-loop (Month 6+): Remediation outcomes as implicit labels

Emerging: LLM-Assisted Labeling

LLMs (GPT-4/Claude/Llama) extract structured RCA labels from ticket text + alarm logs. Early results: 70-85% accuracy, reducing SME effort by 60-80%.

Related Concepts

• Labeled Data Challenge — why this is needed
• GNN+DRL for RCA — bypassing labels entirely
• RCA Dataset Design — structure of a labeled dataset