Construction and Application of Meituan Hotel & Travel Knowledge Graph

Background and Motivation – In Meituan's hotel‑travel vertical search, user needs fall into two categories: precise store search and generic scenario search. Traditional text retrieval struggles with generic scenarios, prompting the construction of a large‑scale knowledge graph using reviews, logs, and other data sources.

Graph Structure – The graph consists of three layers: (1) Taxonomy layer defining label categories, (2) Label layer containing atomic and composite tags with synonym and hierarchical relations, and (3) Supply layer linking tags to POIs, rooms, tickets, and packages. Over 10 million nodes and 100 million edges have been built with ~95% core relation accuracy.

Technical Modules

Knowledge Mining – Extracts tags from structured, semi‑structured (e.g., Baidu Baike) and unstructured data (reviews, logs) using expanded vocabularies, short‑text classification, multi‑task learning, and deep + wide feature fusion.

Relation Discrimination – Identifies synonym and hypernym/hyponym relations via rule‑based similarity, vector indexing, active learning, and statistical co‑occurrence analysis.

Knowledge Association – Aligns mined tags with supply entities using evidence extraction from comments and product pages, followed by multi‑evidence fusion and tree‑model classification.

Applications

Search – Improves recall for composite tag queries, delivering results with relevant scenario tags.

Time‑Sensitive Tag Display – Shows strong‑relevance tags during active periods and only query‑related tags otherwise.

Ranking Lists – Generates popularity lists (e.g., hot springs) based on tag‑driven signals.

Suggestion (sug) – Enhances query suggestions with scenario‑aware tags, reducing irrelevant results.

Risk Control – Classifies tags into high, medium, and low risk, applying filtering and manual review to mitigate sensitive or misleading content.

Summary and Outlook – The graph is built across data, technology, application, and risk layers. Future work includes incorporating multimodal features, expanding graph usage to more filters, and improving risk‑mitigation algorithms to lower operational costs.

Q&A Highlights

Evaluation metrics: node/edge count, accuracy, and downstream coverage.

Tag system evolution: align with business needs and iterative improvement.

Scalability: only high‑frequency tags are indexed for online search.

Multimodal value: combining images with text enhances entity recognition (e.g., identifying roller coasters).