RCA Labeled Dataset Design

Specification for a realistic labeled dataset suitable for training GNN-based RCA models on telecom networks.

Dataset Components

A complete labeled incident requires 6 data artifacts:

Component Format GNN Tensor Purpose
Network topology JSON (nodes + edges) Adjacency matrix edge_index [2×E] Graph structure
Node KPI time series CSV (N nodes × F features × T timestamps) Node feature matrix X [N×F] Per-node state
Edge KPI time series CSV (E edges × D features × T timestamps) Edge attribute matrix edge_attr [E×D] Per-link state
Alarm sequence CSV (chronological) Additional node features or separate graph Fault signals
RCA ground truth JSON (per-node labels) Label vector y [N] Supervision signal
Source ticket JSON (trouble ticket) — Label provenance

Label Schema (5-Class Node Classification)

Label ID Name Definition
0 root_cause The originating faulty element — where the fix should be applied
1 primary_symptom Directly connected to root cause, first-order effect
2 secondary_symptom Affected by cascading propagation, 2+ hops
3 collateral Not in fault path but impacted by traffic rerouting or overload
4 unaffected Normal operation throughout

Alternative target formats:

Temporal Snapshots

Each incident should include multiple timestamps:

Phase What To Capture
Normal baseline All KPIs nominal (reference state)
Early degradation Subtle signals (e.g., rising CRC errors) visible only on root cause element
Fault onset Cascading failures across domains
Peak impact Maximum alarm count, worst KPIs
Post-remediation KPIs recovering to baseline

Node Features (19 dimensions typical)

CPU utilization, memory utilization, connected UEs, DL/UL throughput, PRB utilization (DL/UL), RRC connection success rate, E-RAB setup success rate, handover success rate, CQI average, RSRP average, packet loss rate, latency, active alarm count, interface errors/discards, temperature, power draw.

Plus structural features: node type one-hot (7 types), domain one-hot (3 domains) = 29 total features.

Edge Features (10 dimensions typical)

Utilization, latency, packet loss, bandwidth, CRC errors, input errors, output drops, link flaps, operational status, edge type.

Training Set Diversity

A production training set needs 500–5,000 labeled incidents across fault types:

Class balance: Root cause is ~8% of affected nodes — use focal loss, cost-sensitive weighting, or Graph SMOTE.

Split strategy: By incident (not by node) — 70% train / 15% val / 15% test.

Reference Implementation

A working example dataset with all components is available at knowledge/rca-labeled-dataset/ in this repository. It includes: