Construction and Application of Retail Product Knowledge Graph at Meituan

In the context of new‑retail, Meituan leverages a product knowledge graph as the digital foundation for understanding and reasoning about goods, complementing deep‑learning models with structured, explainable knowledge.

Meituan Brain, launched in 2018, already models billions of entities and triples for merchants; extending this to retail requires handling far more diverse and sparse product data across domains such as grocery, fresh food, and pharmacy.

The graph adopts a five‑level hierarchy (SKU/SPU, Standard Product, Abstract Product, Core Category, Business Category) and captures three perspectives: product attributes (both generic and category‑specific), user cognition (aliases, evaluations, lists), and category relationships (synonyms, hypernyms).

Key construction challenges include low‑quality source data, high dimensionality of attributes, and the need for domain expertise. Solutions involve semi‑supervised and active learning for labeling, BERT‑based NER with graph‑enhanced embeddings for title parsing, and rule‑based, statistical, syntactic, and embedding features for generic knowledge extraction.

Efficiency is improved by decoupling features from tasks, using a unified graph‑mining feature suite, and employing semi‑supervised, active, and distant supervision techniques to reduce manual annotation effort.

For highly specialized domains such as pharmaceuticals, a human‑machine workflow combines expert validation with model‑generated candidates to ensure 100% accuracy for critical knowledge.

Finally, the graph powers downstream applications: structured recall in search, generalized features for ranking models, multi‑modal graph embeddings for recommendation, and both consumer‑facing (filters, tags, explanations) and merchant‑facing (auto‑category and attribute prediction) enhancements.