GPG-VNFE: Towards Foundation Models via Graph Pretraining for Generalizable VNF Embedding

The placement of Virtual Network Functions (VNFs) plays a critical role in the provisioning of services across nextgeneration networks. However, the constantly changing conditions in network topologies, traffic demands and resource availability pose significant challenges to achieving efficient and reliable placement strategies. While Deep Reinforcement Learning (DRL) has been widely explored to address this problem, these approaches have limited generalization capabilities and require retraining when faced with unfamiliar scenarios, such as network topology changes, failure events or evolving service requests (e.g., VNF Forwarding Graphs or VNF-FGs). To address these limitations, we propose GPG-VNFE, a foundation Graph Neural Network (GNN) architecture that integrates graph contrastive learning with generative models as Learnable Priors (GraphCL-LP) within a Transformer-based DRL framework. GPG-VNFE first pretrains two domain-specific GNN encoders using GraphCL-LP, one for substrate network topologies and another for VNF-FG service graphs. Through extensive simulation on a medium-size network topology, we show that GPG-VNFE improves the acceptance ratio of incoming VNF requests by up to 57.45% compared with three baseline state-of-the-art placement approaches. We also show that our proposal lowers the power consumption at the physical layer by up to 11%, outperforming two of the baseline approaches.