CSP Network Data Integration for GNN

A GNN model does not magically “know” a CSP’s network. It is fed a continuously updated graph representation built from existing management systems.

What a GNN Needs

  1. Graph Structure (Topology) — what nodes exist, how they connect, node/edge types
  2. Feature Vectors (State) — per-node KPIs (CPU, throughput, alarms) and per-edge KPIs (utilization, latency, errors)

Both must be continuously updated as the network changes.

Topology Sources

Source Protocol What It Provides Update Frequency
Network Inventory / CMDB REST, TMF639 All managed elements, types, relationships Near real-time
IGP/Routing (OSPF-TE, IS-IS) BGP-LS or controller L3 topology, link metrics Seconds
SDN Controller RESTCONF, OpenFlow Centralized topology view Real-time
NETCONF/YANG RFC 6241 Device config, interface relationships On-demand
3GPP NBI TS 28.532 RAN topology (gNBs, cells, neighbor relations) Real-time
O-RAN E2/O1 E2AP, NETCONF/VES Near-RT RAN topology Real-time
LLDP/CDP L2 discovery Physical adjacency 30-60s

Feature Sources

Source Protocol Features Granularity
PM Counters (3GPP) FTP/Kafka/VES RAN KPIs: RSRP, throughput, PRB utilization 15-min (configurable)
Streaming Telemetry gNMI, gRPC Interface stats, errors, latency Seconds
SNMP SNMPv2c/v3 CPU, memory, temperature 5-min polling
FM Alarms VES, SNMP traps Alarm type, severity, timestamp Immediate
NWDAF (5G) Nnwdaf APIs Network analytics: load, QoS, mobility Real-time

Graph Construction Engine

The critical middle layer that translates raw CSP data into GNN-consumable tensors:

  1. Topology Reconciler — merges sources, resolves conflicts, detects changes
  2. Feature Normalizer — z-score/min-max normalization, handles missing values, encodes categoricals
  3. Graph Builder — creates adjacency matrix A, node feature matrix X, edge feature matrix E

Multi-vendor normalization: Ericsson pmRrcConnEstabSucc and Nokia VS.RRC.ConnEstab.Succ both map to the same rrc_connection_success_rate (float 0-1). Standard models: TMF SID, 3GPP NRM, OpenConfig YANG.

Integration Patterns

Pattern How It Works
Graph Database Hub Neo4j / TigerGraph / Google Spanner Graph stores topology; GNN reads subgraphs. Used by Google Cloud + NetAI.
Streaming Pipeline Kafka → Flink → Feature Store (Redis/Feast) → GNN inference (TorchServe/Triton).
Digital Twin Mirror of production network; GNN trains/infers on twin data; safe for what-if simulation.