Graph Neural Networks: The New AI Darling and Its Industrial Deployment Challenges

Due to the limitations of deep learning in interpretability and handling non‑regular data, graph neural networks (GNNs) that combine graph computing with deep learning have become a hot research direction in both academia and industry.

At the KDD 2019 conference, Ant Financial hosted a session on GNN research and applications, where InfoQ interviewed Ant Financial AI researcher Song Le to discuss the practical use of deep learning and GNNs in large‑scale industrial scenarios, current challenges, and future directions.

GNNs are now regarded as the "new AI darling" because graph structures offer strong expressive power, and GNNs provide better performance and interpretability, leading to a surge of frameworks and tools that accelerate the field.

GNNs enable graph representation learning or graph embedding, allowing supervised, semi‑supervised, and reinforcement learning on graph data; they transform nodes, edges, or sub‑graphs into vectors whose distances preserve original graph similarity, bridging traditional graph analysis with machine‑learning methods in recommendation systems, knowledge‑graph construction, and more.

By addressing deep learning’s difficulty with irregular data, GNNs expand the applicability of neural networks and improve model interpretability, making them valuable for business scenarios such as recommendation, disambiguation, and fraud detection; Ant Financial has already deployed GNNs widely in inclusive finance recommendation and risk control.

Song Le highlighted two intriguing new applications: (1) reasoning over massive knowledge graphs that link millions of registered enterprises, and (2) dynamic graphs that incorporate temporal transaction data for real‑time loan‑approval models.

He estimates that GNNs have been applied in dozens of Alibaba business lines, though the technology is still in its early stages.

The main obstacles to large‑scale industrial deployment are the need for distributed graph‑computing platforms capable of handling billions of nodes and edges, and the current lack of mature open‑source solutions that can efficiently support such scale.

Training GNNs at scale requires tight interaction with distributed graph storage, frequent random access to nodes and neighborhoods, and efficient data movement; Ant Financial’s recent distributed graph storage and training platform has reduced training time for billion‑node graphs from days to under an hour.

Online inference is also challenging because GNN embeddings are not computed in real time; updating embeddings instantly as new edges appear (e.g., financial transactions) while meeting strict latency requirements demands close cooperation with underlying system architectures.

Overall, graph computation itself is difficult due to irregular operations, and adding deep‑learning layers further increases complexity; achieving fast training and prediction is essential for engineers to experiment with different GNN architectures.

Both academia and industry view these issues as frontier problems, and despite efforts from Google, Facebook, and others, no dominant open‑source platform for large‑scale distributed GNN computation has yet emerged.

For more details, readers are invited to click the "Read Original" link to visit Ant Financial’s official website.