Construction and Application of Retail Product Knowledge Graph at Meituan

The article introduces the retail product knowledge graph as a foundational digital asset for Meituan's new‑retail business, highlighting its role in providing precise, structured understanding of billions of products, merchants, and user comments.

Background : Knowledge graphs, together with deep learning, are key drivers of AI. Meituan’s “Meituan Brain” has built a large‑scale knowledge graph covering tens of billions of entities and triples across food, hotel, finance, and other domains.

Retail Exploration : To support services such as Meituan Flash, Meituan Buy Vegetables, and Meituan Preferred, a dedicated retail product knowledge graph is required to model products, attributes, user perceptions, and scenarios.

Challenges : (1) Low‑quality information sources (sparse titles/images); (2) High dimensionality of product attributes (hundreds of dimensions across categories); (3) Dependence on common sense and professional domain knowledge.

Hierarchical System (L1‑L5) : L1 – SKU/SPU (transactional unit) L2 – Standard product (objective facts, e.g., brand, flavor) L3 – Abstract product (series, user‑perceived concepts) L4 – Core category (e.g., egg, strawberry cream) L5 – Business category (front‑end taxonomy defined per business stage).

Multi‑Dimensional Attribute Construction : Attributes are captured from three sources – titles, images, and semi‑structured data. Title parsing is treated as a sequence labeling task using NER enhanced with graph‑based entity information, relational links, and node‑type embeddings to inject common‑sense knowledge.

Efficiency Improvements : Knowledge extraction is modeled as three classification tasks (node, relation, association). A unified feature framework combines rule‑based, statistical, syntactic, and embedding features, while semi‑supervised learning, active learning, and distant supervision reduce labeling effort.

Human‑Machine Collaboration for Professional Domains : In the pharmaceutical sector, weak professional knowledge is mined from manuals and encyclopedias with expert‑guided rules, whereas strong professional knowledge relies on model‑generated candidates reviewed by experts to ensure 100% accuracy.

Applications : • Structured recall – leveraging graph‑based semantics for ambiguous queries. • Ranking model generalization – using multi‑granular category and attribute features, as well as graph embeddings. • Multi‑modal graph embedding – integrating images, titles, and textual descriptions via MKG Entity Encoder and MKG Attention Layer. • User‑side optimization – providing filter options, tags, and recommendation reasons. • Merchant‑side optimization – real‑time category prediction and attribute extraction from product titles.

The article also includes author information, recruitment notices, and links to related Meituan research papers.