Graph Deep Learning: Methods, Frameworks, and Industrial Applications

In recent years, graph deep learning has attracted significant attention in both academia and industry, leading to numerous research works and practical applications.

This article introduces the emergence and development of graph deep learning, and presents Alibaba's Euler graph deep learning framework and its use in search advertising as a case study.

Deep learning excels at processing large-scale data, and graph neural networks (GNNs) extend this capability to graph-structured data, enabling relational reasoning and improved interpretability.

Traditional deep models such as MLP and CNN can handle regular graph-like structures (e.g., images, text), but real-world graphs are irregular and require specialized GNN architectures.

Two main families of GNN methods exist: spatial-based approaches that aggregate neighbor information, and spectral-based approaches that operate in the graph Fourier domain. The seminal GCN model unified these perspectives.

Representative GNN models include GCN, GraphLSTM, GAT, GraphSAGE, FastGCN, and MixHop, each addressing different aggregation or sampling strategies.

While many deep learning frameworks (TensorFlow, PyTorch) support general models, dedicated graph learning frameworks are less mature. Euler, open‑sourced by Alibaba in 2019, provides a distributed engine and algorithm layer for large‑scale graph training.

Euler supports billions of nodes and edges, heterogeneous graphs, and offers flexible sharding and replication to achieve high throughput.

In Alibaba's search advertising system, a heterogeneous graph of queries, items, and ads is constructed. Three types of edges—click, text similarity, and shared bidding—capture user behavior.

The SMAD (Scalable Multi‑view Ad Retrieval) pipeline consists of large‑scale graph construction, parallel DNN training via random walks on the graph, and online inference using approximate nearest‑neighbor search on learned embeddings.

Experiments show that the Euler‑based solution improves click‑through rate by over 1.5% compared with baseline methods.

The article concludes that graph deep learning remains a vibrant research area with many open challenges, such as online inference and advanced graph representations.