Title: Theoretical Bounds on the Transferability of Graph Neural Networks in Wireless Conflict Graphs

Abstract

Graph Neural Networks (GNNs) have become a potent instrument for wireless resource allocation by capitalizing on the inherent graph topologies of communication systems. A key advantage of these models is their transferability, allowing algorithms trained on compact graphs to scale to extensive deployments with minimal degradation in performance—a critical feature for expanding modern networks. Given that wireless environments are typically sparse, with each node linking to only a limited subset of peers, this study investigates the theoretical foundations of GNN transferability within graphs generated from sparse Random Geometric Graphs (RGGs). Specifically, the research concentrates on the conflict graphs of RGGs, which are utilized to represent interference patterns between transmission links. By analyzing the proximity between RGGs and Deterministic Grid Graphs (DGG), we derive bounds on the performance penalty incurred when transferring models across different scales. These theoretical insights are empirically validated through link scheduling tasks, where our learned strategies consistently surpass established benchmarks in large-scale scenarios. Lastly, the paper assesses how the underlying theoretical assumptions influence actual performance outcomes.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC