Graph Neural Network Based Scene Modeling for Food Delivery CTR Prediction

Introduction

Meituan Waimai's advertising CTR prediction faces a highly scenario‑driven problem: user decisions are influenced by geographic location (LBS), merchant status, and time slots, creating a short‑term, context‑rich decision chain. Traditional feature‑cross or sequence models struggle to capture these nuances.

Challenges

Feature crossing introduces noise and cannot fully represent scene interactions.

Sequence models depend on abundant user behavior and often trap high‑frequency users in narrow interest circles while under‑representing low‑frequency users.

Both paradigms lack a complete, high‑order representation of the user‑scene‑item relationship.

Why Graph Technology

Graph neural networks (GNNs) provide structured, relational, global, and highly generalizable representations, making them suitable for modeling the complex, heterogeneous interactions in food‑delivery scenarios. The article positions graph modeling as the next step after manual features and NN‑based sequence models.

Scenario Feature Crossing via Graphs

Features are treated as nodes in a bipartite graph; edges encode potential cross‑interactions. A parameter matrix M (randomly initialized) stores node embeddings; L0 regularization sparsifies M to reduce noisy crosses. The workflow includes:

Initialize M and compute dot‑product edge scores.

Apply L0 penalty to push irrelevant edges toward zero.

Binarize M to select active cross pairs.

Aggregate neighbor features with an attention mechanism.

Use the aggregated vectors as updated feature embeddings for downstream CTR training.

This graph‑based crossing yields more informative, noise‑reduced cross features.

Behavior Subgraph Expansion

To break the interest‑closure of sequence models, user sessions are transformed into POI subgraphs. Within a session, clicked POIs form a connected graph; overlapping POIs across sessions link these graphs into a global behavior graph. For each POI, one‑hop and two‑hop neighbors are retrieved, and a subgraph convolution aggregates their embeddings, producing a richer POI representation that incorporates both personal and global context.

Three aggregation strategies are employed:

Attention‑weighted aggregation of neighboring POI embeddings.

Context‑aware weighting to adapt to different user sessions.

Joint aggregation with the target POI embedding during training.

Metapath‑Based Scene Graph Modeling

Events are abstracted as User‑POI‑Context triples; chains of such events form heterogeneous scene graphs. Metapaths (e.g., U‑C‑P‑C‑U) capture multi‑step relationships. The modeling steps are:

Encode user and merchant attributes via GraphSAGE and attach them to node embeddings.

From a given user, enumerate multiple metapath instances (event chains) and retrieve their vector representations.

Concatenate these vectors and apply attention with the target user to weigh each metapath's relevance for the CTR task.

Engineering Deployment

The final CTR model integrates three plug‑in modules—feature‑graph crossing, subgraph expansion, and metapath scene graph—following low‑coupling, high‑cohesion principles. To meet latency constraints, long‑term behavior subgraphs are pre‑computed offline, stored in Redis, and queried online, eliminating costly real‑time convolutions. The serving pipeline consists of:

End‑to‑end training with joint graph modules.

Offline subgraph embedding aggregation and Redis ingestion.

Model reconstruction that replaces runtime subgraph computation with Redis look‑ups.

Online experiments show that the added latency is negligible (TP99/TP999 unchanged).

Effectiveness

Offline AUC gains:

Subgraph expansion + joint training: +2 ‰.

Feature‑graph crossing module: +1 ‰.

Scene‑graph interaction: +2.5 ‰.

Both high‑frequency and low‑frequency users see CTR and GMV improvements, and exposure per category increases, especially for high‑frequency users. Time‑of‑day analysis shows stronger tea‑time exposure for milk‑tea categories, while spatial analysis reveals higher weekend exposure for convenience‑store items, confirming the model's ability to capture temporal and spatial scene signals.

Conclusion and Outlook

Graph‑based scenario modeling effectively addresses the unique challenges of food‑delivery recommendation, delivering measurable gains without sacrificing serving efficiency. Remaining issues include computational limits on exhaustive feature crossing and the reliance on manually defined metapaths; future work will explore automated metapath discovery and broader feature‑graph scalability.