Low-Latency Relay Selection in NR-V2X Vehicular Communications via Graph Isomorphism Networks with Edge Features

In dense urban environments, vehicle-to-infrastructure (V2I) links often suffer from blockages and rapid channel variations. While multi-hop relaying can restore connectivity, selecting optimal relay paths is an NP-hard problem that typically requires computationally expensive Mixed-Integer Linear Programming (MILP). The authors investigate whether a Learning-to-Optimise (L2O) framework can provide near-optimal relay selection with the low latency required for real-time NR-V2X communications.

The researchers model V2X snapshots as directed graphs where nodes represent vehicles and infrastructure, and edges represent potential wireless links. They employ a Graph Isomorphism Network with Edge Features (GINE) to process these graphs. The GINE architecture is specifically designed to incorporate link-level radio attributes into the message-passing process. The model is trained using ground-truth configurations generated by an offline MILP oracle. Additionally, the authors introduce a hybrid strategy, GP-MILP, which uses the GINE model's output to prune the search space for the MILP solver, effectively bridging the gap between fast neural inference and exact mathematical optimization.

The GINE model demonstrates high accuracy (0.9589) and F1-score (0.9544) in matching the decisions of the MILP oracle at the link level. In terms of network performance, the framework provides significant end-to-end connectivity gains—up to 12% compared to a 1-hop baseline—while keeping inference latency strictly under 5 ms. The hybrid GP-MILP approach successfully maintains MILP-equivalent solutions while reducing solver runtimes to below 30 ms for over 98% of tested instances, making it feasible for real-time deployment.

This work addresses the critical trade-off between the optimality of traditional optimization and the speed of machine learning in vehicular networks. By demonstrating that GNNs can effectively guide or replace expensive solvers, the paper provides a scalable path toward reliable, low-latency connectivity in future 5G-Advanced and 6G V2X systems.