GraphVNE: Graph-Level Matching for Efficient Virtual Network Embedding in Edge Computing

doi:10.1109/JIOT.2026.3656629

Return

GraphVNE: Graph-Level Matching for Efficient Virtual Network Embedding in Edge Computing

2026-01-21

PRE

DOI：10.1109/JIOT.2026.3656629

Original

Original request for help

Save

Abstract

En 中文

Virtual network embedding (VNE) plays a crucial role in network virtualization, particularly within edge computing environments where low-latency and efficient resource allocation are required under dynamic workloads and fragmented resources. Although existing VNE approaches have evolved to better adapt to dynamic demands, they often lead to severe resource fragmentation, limiting overall efficiency. To address this issue, we propose GraphVNE, a reinforcement learning (RL)-based framework that incorporates graph-level matching information to obtain more structurally aware representations and thereby enhance VNE performance. Specifically, we design a graph matching module (GMM) that computes node-to-graph matching scores between virtual and physical networks, which are then used to enrich feature representations and guide more informed decisions in the embedding policy. To further integrate heterogeneous features, we design a feature fusion method that combines inner-graph features with cross-graph matching features, producing two enriched representations optimized for bidirectional action selection. Extensive experiments on the GEANT dataset show that our design consistently outperforms state-of-the-art approaches, with an exactly 13.1% increase in revenue-to-cost (R2C) ratio over the strongest approach. It also demonstrates robust performance in high request-density scenarios, supporting more reliable and scalable edge services. The results demonstrate GraphVNE’s potential to significantly improve VNE efficiency, offering a scalable and effective solution for resource-constrained edge environments.

Keywords:

Edge computing

reinforcement learning (RL)

virtual network embedding (VNE)