GATNextHop Uses GNNs for Cross-Topology Shortest Path Routing

Traditional shortest-path algorithms, such as Dijkstra's, are highly accurate but require recomputation for every change in network topology, posing scalability challenges in dynamic or large-scale networks. This paper introduces GATNextHop, a novel model that leverages a Graph Attention Network (GAT) to approximate shortest path routing. The core objective is to determine if a GNN can learn routing heuristics that generalize effectively across diverse network structures. The GATNextHop model is trained using synthetic graphs and then rigorously evaluated on real-world Internet Service Provider (ISP) networks sourced from the Internet Topology Zoo. The research aims to quantify the trade-offs between the learned approach of GATNextHop and classical routing algorithms like Dijkstra's, focusing on key performance indicators such as accuracy, inference speed, and generalization capabilities. This work explores the potential for GNNs to offer a more scalable and adaptive routing solution.