Telecom GNN RCA Labeled Dataset

A realistic labeled dataset for training Graph Neural Networks to perform Root Cause Analysis on telecom network faults. Based on a backhaul SFP failure causing cascading degradation across RAN, transport, and core domains.

Quick Start

source .venv/bin/activate
python scripts/07_pyg_graph_construction.py

Graph Structure

• Network Topology - 12-node heterogeneous graph (gNBs, routers, core NFs) with 14 edges across 3 domains

Node & Edge Features (Time Series)

• Node Features - 19 KPIs per node at 4 temporal snapshots (baseline → fault → recovery)
• Edge Features - 9 KPIs per edge at 4 temporal snapshots, showing CRC error escalation on fault link

Fault Data

• Alarm Sequence - 47-alarm storm across 10 network elements over ~47 minutes
• Trouble Ticket - Original ServiceNow ticket INC00847291 with work notes from 4 engineers

Labels & Training

• RCA Ground Truth Labels - Per-node 5-class labels (root_cause, primary/secondary symptom, collateral, unaffected) with propagation graph
• Training Batch - 8 diverse incident types for batch training with class balance guidance

Code

• Graph Construction Script - Converts raw data to PyTorch Geometric Data object (node/edge feature matrices + labels)

Data & Code Directories

• samples/ - Raw data files (topology JSON, KPI CSVs, alarm sequence, labels, trouble ticket, training batch)
• scripts/ - Graph construction Python script