Large-Scale Heterogeneous Graph Modeling and GraphET Engine for Meituan Food Delivery Search Advertising

This article presents the application of graph technology to improve the weak-supply filling of Meituan food delivery search advertising, thereby increasing traffic monetization efficiency. It introduces the evolution of heterogeneous large‑graph modeling for multiple scenarios and the construction of the GraphET large‑scale graph training and online inference engine, which supports billions of nodes and edges.

1. Introduction Meituan food delivery’s ad system relies on a recall stage to retrieve high‑quality candidates from massive item pools. Due to LBS constraints, the recall system is divided into no‑supply, high‑supply, and weak‑supply regions. Weak‑supply filling is handled by a search‑recommendation pipeline that matches user intent with available supply in real time.

2. Graph Technology and Engine Overview Recent advances in graph neural networks (GNNs) are surveyed, covering the transition from unsupervised walk‑based methods to message‑passing paradigms, graph pre‑training, and dynamic graph learning. Industrial requirements for large‑scale graph training (hundreds of billions of edges) and low‑latency online inference are discussed, highlighting the limitations of existing open‑source frameworks.

3. Evolution of Heterogeneous Large Graph in Search Recommendation The authors propose a multi‑scenario heterogeneous graph that evolves from single‑scenario fine‑grained modeling to unified pre‑training + downstream fine‑tuning, and finally to GPT‑enhanced pre‑training with prompt‑based fine‑tuning. Specific techniques include:

EM‑based language‑enhanced denoising graphs for single‑intent noise reduction, achieving +3.7% offline recall.

Multi‑intent contrastive learning that distinguishes intent‑specific node representations, yielding +1.8% recall for the contrastive component and +3.8% average recall across tasks.

WM (Weak‑Supply Multi‑scenario) large‑graph construction using user profiles, session sequences, and item nodes, resulting in a graph with billions of nodes and hundreds of billions of edges.

Unified large‑graph pre‑training with shared bottom layers and heterogeneous top layers, improving multi‑task average recall by +4%.

Generative‑enhanced retrieval (GAR) combined with GNNs, delivering +1% multi‑task recall and +10% zero‑shot recall after soft‑prompt fine‑tuning.

4. Large‑Scale Graph Engine GraphET Construction Built on DGL v0.7, GraphET supports:

Training pipelines that handle graph construction, sampling, aggregation, and end‑to‑end modeling on GPUs.

A worker‑process/parameter‑server architecture that stores graph structures in shared memory and embeddings in a multi‑level storage hierarchy (GPU, RAM, SSD).

Optimizations such as SSD batch read aggregation, object‑pool allocation, and asynchronous garbage collection.

The system also includes a multi‑process online inference framework to overcome Python GIL limitations and GPU memory constraints. Model loading is split between dense parameters (GPU) and large embedding tables (CPU memory), enabling many parallel inference workers.

5. Conclusion and Outlook Graph neural networks demonstrate strong potential for search advertising. The paper summarizes the deployment of a heterogeneous large‑graph solution for Meituan’s food‑delivery ad scenario and outlines future directions, including GPT‑style universal graph models, trillion‑edge graph construction, and further acceleration of online graph inference.