Applying Knowledge Graphs to Meituan's Recommendation System

Meituan, a large lifestyle service platform, constructs a massive knowledge graph called "Meituan Brain" that contains billions of entities such as merchants, products, dishes, and user reviews, enabling rich semantic understanding of user queries and merchant attributes.

The recommendation scenario is location‑based (LBS) and differs from e‑commerce or video recommendation; it faces strong explainability demands, diverse domains, data sparsity, and complex spatio‑temporal contexts.

Knowledge graph applications are divided into explicit (structured information display, path‑based recall, graph‑driven recommendation reasons) and implicit (embedding the graph into downstream models) to address these challenges.

For explicit use, the system displays structured tags (e.g., "large table", "private room") and uses graph paths to suggest related queries, improving click‑through rates and reducing no‑result cases.

To model multi‑interest users across many domains, a dual‑memory network based on Neural Turing Machines is proposed, with separate memories for item interests (MI) and user interests (MU). Item and entity embeddings are first generated by a heterogeneous graph neural network (HGNN), then processed by the dual memory to produce fine‑grained interest distributions.

Extensive offline experiments on Meituan's review business show the dual‑memory model outperforms single‑interest and existing multi‑interest baselines on all metrics; ablation studies confirm the importance of both item and user memories and the KL‑divergence regularizer.

Cross‑domain learning leverages the rich graph and high‑traffic domains (e.g., food delivery) to enhance sparse domains (e.g., grocery, leisure) by constructing a larger heterogeneous graph that includes external search behavior, yielding significant gains in click‑NDCG.

The conclusion highlights the critical role of knowledge graphs in both explicit guidance and implicit representation, and outlines future work on multi‑interest modeling, multimodal items, spatio‑temporal interactions, recommendation fairness, cross‑domain learning, and graph pre‑training.

A Q&A session addresses practical questions about graph‑based recall versus ranking models, template‑driven recommendation reason generation, and online performance of the dual‑memory network.