Graph Isomorphism Networks Cut V2X Relay Latency to Under 5ms

Reliable low-latency uplink in dense urban C-V2X networks is critical for autonomous driving, but direct vehicle-to-infrastructure links often degrade due to blockages. Multi-hop relaying solves coverage issues, but optimal relay selection under radio, capacity, and routing constraints is NP-hard — traditionally solved via Mixed-Integer Linear Programming (MILP), which scales poorly with graph size. The paper proposes an edge-aware Learning-to-Optimise framework: each V2X snapshot is modeled as a directed graph with node features (vehicle state, traffic demand) and edge features (radio-link capacity). An offline MILP oracle supervises a Graph Isomorphism Network with Edge Features (GINE), enabling edge-level relay activation in a single forward pass with tightly bounded inference latency.

The experiments use a large-scale dataset from an OSM-SUMO-GEMV$^2$ pipeline. GINE closely matches MILP decisions at the link level (accuracy 0.9589, F1-score 0.9544 on validation) and yields consistent end-to-end connectivity gains — up to 9.2% with four RSUs and 12% with two RSUs over a 1-hop MILP baseline. Inference latency remains under 5ms for all evaluated instances. The hybrid GP-MILP strategy prunes the MILP search space using GINE predictions, preserving equivalent objective values while reducing solver runtime below 30ms for over 98% of graph instances — making MILP-grade optimization compatible with stringent NR-V2X latency budgets.