Construction and Application of Retail Product Knowledge Graph at Meituan

In the context of new‑retail, Meituan builds a product knowledge graph as a digital foundation that provides a three‑dimensional, intelligent, and commonsense understanding of goods, enabling downstream business scenarios such as search and recommendation.

The "Meituan Brain" knowledge graph, launched in 2018, already models billions of entities and triples across food, hospitality, finance and other domains, forming a large‑scale "knowledge brain" for life‑service services.

Meituan's new‑retail expansion (flash‑sale, grocery, selection, etc.) requires a more granular product graph because product data are sparse, diverse and need commonsense reasoning; thus a five‑level hierarchy (SKU/SPU, Standard Product, Abstract Product, Subject Category, Business Category) is defined.

Multi‑dimensional modeling covers product attributes (both generic and category‑specific), user cognition (aliases, evaluations, lists) and category perspectives, while cross‑business alignment links objective knowledge to front‑end categories and back‑end SKUs/SPUs.

Key challenges include low‑quality source data, high dimensionality (hundreds of attributes), and the need for both common‑sense and professional domain knowledge; solutions involve semi‑supervised learning, active learning, distant supervision, and expert‑in‑the‑loop verification.

Construction pipelines include hierarchical taxonomy creation, category labeling via NER and BERT‑based matching, standard/abstract product linking, attribute extraction from titles, images and semi‑structured data, and a unified feature system (rule‑based, statistical, syntactic, embedding) to improve efficiency.

Human‑machine collaboration is applied especially in the pharmaceutical domain, separating weak and strong professional knowledge and combining model predictions with expert review.

Applications of the product graph span structured recall, ranking model generalization, multi‑modal graph embeddings, and C‑side (filters, tags, recommendations) and B‑side (automatic category and attribute prediction) optimizations, ultimately enhancing user experience and merchant efficiency.