Edge‑Cloud Perspectives on Graph Neural Network‑Based Recommendation Systems

This article provides an overview of graph neural network (GNN) recommendation systems from an edge‑cloud perspective, outlining four key aspects of their feasibility on the device side.

1. Evolution of Underlying Compute: Over the past two decades, computing has shifted from cloud‑centric models to a dual paradigm that includes edge computing. This shift enables large‑scale AI model training in the cloud while allowing lightweight AI inference on devices, creating a polarised development of AI.

2. Personalization on the Device Side: The edge‑cloud view is likened to a global graph versus localized sub‑graphs. Global graphs offer comprehensive relational data and strong generalization, whereas local sub‑graphs capture individual user behavior for fine‑grained personalization. Combining both improves recommendation performance, as demonstrated by the Ada‑GNN model presented at WSDM 2022.

3. Edge‑Cloud Collaborative Implementation: Three deployment modes are discussed: (a) session‑based recommendation in the cloud, (b) single‑item recommendation, and (c) personalized edge models. A meta‑controller can dynamically select the appropriate mode. Techniques such as causal‑effect estimation with proxy models enable the construction of synthetic datasets for training the controller.

4. Security Concerns on the Device Side: Deploying personalized GNN models on devices introduces attack vectors such as evasion, data poisoning, and backdoor attacks, raising significant security risks for recommendation systems.

The article also addresses practical challenges like model compression (pruning, quantization, distillation, and knowledge‑distillation via adversarial metrics), split deployment of GNN layers between cloud and edge, handling distribution shift, and online personalization using parameter‑efficient transfer learning. Real‑world experiments from Alibaba and Huawei illustrate the benefits of split deployment, achieving kilobyte‑level model sizes on devices.

A brief Q&A clarifies that sub‑graph transmission overhead is minimal because only small, neighbor‑level sub‑graphs are sent, and that the approach balances accuracy with resource constraints.