Construction and Application of Meituan's On‑site Comprehensive Knowledge Graph

Meituan's on‑site comprehensive business covers a wide range of local life services such as leisure, beauty, parenting, wedding, and pets. To improve supply‑demand matching, Meituan built a user‑centric knowledge graph (GENE – General Needs net) that links user needs with merchants, products, and content.

Knowledge‑graph architecture consists of six layers: scene‑need layer, scene‑element layer, concrete‑need layer, need‑object layer, industry‑system layer, and supply layer. Each layer captures different granularity of user demand and connects it to corresponding supply.

Construction challenges include handling diverse user needs, the complexity of hundreds of local‑life industries, and the high quality requirements of the graph. Solutions involve multi‑source, multi‑modal data mining, hierarchical node extraction, and joint modeling of statistical and semantic features.

Key construction steps :

Design of the six‑layer schema and definition of scene‑need, scene‑element, concrete‑need, need‑object, industry, and supply nodes.

Node mining (demand, objects, attributes) using pipelines of keyword extraction → clustering → dimension refinement, combined with unsupervised expansion and supervised BERT‑CRF labeling.

Relation building (synonym, hierarchical, attribute) via pattern extraction and BERT‑based entity‑relation models.

Supply association through multi‑stage entity linking (recall, ranking, aggregation) using Wide&Deep, BERT, and multimodal (text + image) matching.

Industry‑system construction by reusing existing category trees, adding multi‑dimensional attributes, and training a multimodal classifier to map merchants and products to categories.

Data outcomes : after more than a year, the graph covers 60+ industries, contains over 400 k demand nodes, billions of relations, and maintains >90 % precision‑recall.

Practical applications :

Search : enriches recall and explainability, e.g., medical‑beauty queries are matched to relevant services.

Recommendation : improves recall and ranking by injecting graph features into recommendation models across Meituan’s homepage and channel pages.

Intelligent information display : aggregates supply into demand‑driven icons, provides tag filters, and generates recommendation reasons from demand‑supply links.

Case study – Script‑kill industry : the graph was extended to standardize script‑kill supply, using rule‑based, semantic, and multimodal aggregation to build a standard script library, then linking merchants and content to standardized scripts, resulting in new category launch, recommendation gains, and clearer information presentation.

Future outlook : expand the graph to user nodes, cover the full decision‑making chain (including fulfillment), continuously enrich industry coverage, and deepen graph‑based applications through improved representation learning.

Q&A : discussed template‑based triple extraction, industry expansion timelines, and offline vs. online label recall mechanisms.