Edge‑Cloud Collaborative Graph Neural Network Recommendation Systems: Architecture, Personalization, Model Compression, and Security

Recent advances in Graph Neural Networks (GNNs) have spurred research on their use in recommendation systems. The article first outlines how compute paradigms have shifted from cloud‑centric to edge‑centric architectures, emphasizing the balance between large‑scale training in the cloud and lightweight inference on devices.

It then examines personalization on the edge by comparing global graphs (cloud) with local sub‑graphs (device). Techniques such as Ada‑GNN, which combine whole‑graph modeling with sub‑graph adaptations, are presented as ways to improve individualized recommendations.

The discussion moves to model compression for edge deployment, covering pruning, quantization, knowledge distillation, and split‑deployment strategies that keep shared base layers in the cloud while placing only the GNN layers on the device, thereby reducing storage and compute overhead.

Implementation of cloud‑edge collaborative recommendation is described through three serving modes—session recommendation, single‑item recommendation, and edge‑side personalized models—and a meta‑controller is proposed to dynamically select the appropriate mode based on user behavior and data sparsity.

Finally, the article highlights security risks inherent to edge‑side GNN deployment, such as evasion, data poisoning, and backdoor attacks, stressing the need for robust defenses when exposing personalized models to open environments.