Building and Scaling Ant Group's Merchant Knowledge Graph: Architecture, Construction, Fusion, and Open Platform

Ant Group has accumulated massive merchant data across offline stores, online mini‑programs, and brand interactions, and faces challenges of sparse basic information, heterogeneous multi‑source data, and the need for a rich, precise, and efficient merchant knowledge infrastructure.

Merchant Knowledge Graph Overview : The graph connects merchants and users, integrating offline POS data, online mini‑programs, and various third‑party channels, providing a unified view of merchant types, locations, services, and their relationships to user behaviors.

Construction & Fusion :

Data sources include structured business data (ODPS, MySQL), semi‑structured info‑boxes, and unstructured text/images.

Knowledge modeling defines a schema with concept, entity, and event domains, capturing abstract concepts, concrete merchant entities, and user behavior events.

The processing pipeline (DAG) performs data ingestion, cleaning, entity/relationship extraction, schema mapping, entity linking and normalization, quality inspection, and storage.

Storage is layered to support real‑time queries, platform operations, and large‑scale graph computation.

Knowledge operation enables query, editing, visualization, and large‑scale graph embedding and reasoning.

Cross‑graph fusion combines merchant graphs with other domain graphs (e.g., credit, security) by aligning identical entities and linking different entity types, achieving knowledge reuse, reducing data duplication, and accelerating business value.

Cognitive Applications : A dual‑tower vector recall model incorporates dynamic graph embeddings (using a proprietary delta‑graph encoder) to improve coupon recommendation, mitigating popularity bias and enhancing coverage of long‑tail items.

Open Platform :

Open knowledge releases concept‑level data such as intents, brands, categories, and tags.

Three public tasks are released: multi‑intent classification for mini‑programs, hierarchical concept prediction, and brand entity recognition.

SDKs (Java and Python) provide data integration, graph storage access, schema management, and algorithm deployment, enabling developers to build, query, and reason over graphs easily.

In summary, Ant Group's merchant knowledge graph now contains over 30 billion entities and 400 billion relationships, and the team plans to progressively open the data and tooling to the community to foster further research and application.