Tulong: An Industrial Graph Neural Network Framework and Learning Platform at Meituan

Meituan Search and NLP teams have built a graph neural network (GNN) framework called Tulong together with a supporting graph learning platform, aiming to improve model scale, iteration efficiency, and ease of use for large‑scale industrial applications.

The article first outlines the motivation: graphs naturally model relationships in many domains (social networks, e‑commerce, knowledge graphs), but existing deep‑learning frameworks (TensorFlow, PyTorch) perform poorly on sparse graph data, and open‑source GNN libraries (PyG, DGL) still have limitations in scalability, configurability, and business integration.

Four main challenges are identified: (1) comprehensive support for heterogeneous GNN models; (2) cost‑effective training on billion‑edge graphs; (3) seamless integration with business pipelines; (4) developer‑friendly yet extensible design.

To address these, Tulong introduces a three‑layer system: a graph and deep‑learning engine (MTGraph) that provides compact graph storage, fast neighbor sampling, and compatibility with PyTorch/DGL; the Tulong framework that encapsulates common GNN components (message, aggregation, update functions) and offers fine‑grained configuration for homogeneous, heterogeneous, and dynamic graphs; and a graph learning platform that supplies visual tools for data ingestion, graph construction, experiment management, and automated workflow orchestration.

The model framework abstracts GNN computation into four operators—message, node‑level aggregation, edge‑type aggregation, and update—covering homogeneous, heterogeneous, and dynamic graph paradigms. Configuration files allow users to assemble these operators into popular models such as GCN, GraphSAGE, JKNet, RGCN, and temporal GNNs without writing custom code.

Performance optimizations focus on two aspects. First, a compressed graph data structure reduces memory consumption dramatically (e.g., training the MAG240M‑LSC dataset on a single machine with only 15 GB instead of >100 GB). Second, sub‑graph sampling is accelerated by a lightweight random number generator, probability quantization, and timestamp indexing, achieving 2–4× speedups over DGL on both static and dynamic graphs.

Empirical results show that Tulong can train billion‑edge graphs on a single machine (e.g., 2 billion‑edge graph in 0.5 h/epoch) and cut memory usage by up to 60 % compared with existing frameworks.

Finally, the platform streamlines the end‑to‑end workflow: dataset management (Spark‑based graph construction), experiment management (configurable training templates), and process automation (pipeline scheduling, model export). The authors conclude that Tulong has successfully powered search, recommendation, advertising, and delivery services at Meituan and that further research and engineering efforts will continue to expand its capabilities.