Applying Knowledge Graphs to Meituan's Recommendation System: Architecture, Challenges, and Future Directions

Meituan, a leading life‑service platform, has built a massive knowledge graph called "Meituan Brain" that covers millions of merchants, billions of items, and tens of billions of user reviews, providing rich semantic information for both user intent understanding and merchant attribute extraction.

The recommendation scenario is location‑based (LBS) and differs from e‑commerce or video recommendation; it faces four main challenges: strong explainability requirements, a wide variety of domains (food, travel, entertainment, etc.), data sparsity in many sub‑domains, and complex spatiotemporal contexts.

Knowledge graph applications are divided into explicit and implicit uses. Explicitly, structured information (e.g., "large table", "private room") is displayed to users and graph‑based paths are used for recall and to generate personalized recommendation reasons. Implicitly, the graph is embedded into downstream models to enrich item and user representations.

For multi‑interest modeling, a dual‑memory network built on a Neural Turing Machine (NTM) is proposed. Heterogeneous Graph Neural Networks (HGNN) first encode items and linked entities, then two memory modules store item‑level interest distributions (MI) and user‑level interest distributions (MU). Read/write operations enable fine‑grained interest extraction and aggregation.

Extensive offline experiments on Meituan's review business show that the dual‑memory model outperforms single‑interest and existing multi‑interest baselines across all metrics. Ablation studies confirm the importance of both item and user memories as well as KL‑divergence regularization.

To address data sparsity, cross‑domain learning leverages high‑traffic domains (e.g., food delivery) and the shared knowledge graph to enhance low‑traffic domains (e.g., grocery, leisure). Temporal sampling and heterogeneous graph construction further improve representation quality, yielding significant gains in click‑NDCG.

The authors outline future work: refining multi‑interest and multimodal item modeling, deeper exploration of spatiotemporal user behavior, and advancing recommendation fairness, cross‑domain transfer, and graph pre‑training.