Intelligent Training Powered by Knowledge Graphs: Progress and Applications at Beike Real Estate

Speaker: Sun Baqun, Senior Technical Manager at Beike Real Estate

Introduction: Sun Baqun introduces the concept of intelligent training based on knowledge graphs, outlining five key topics: entity graph progress, reasoning graph progress, applications of knowledge graphs in intelligent training, the establishment of a knowledge operation loop, and future thoughts on the industry graph.

1. Entity Graph Progress

Beike aims to build a knowledge middle‑platform through four steps—knowledge ingestion, organization, application, and operation—to become the standard‑setter for real‑estate industry knowledge. Since 2018, massive data has been ingested, cleaned, and fused into a layered knowledge base that supports business intelligence across multiple departments.

2. Knowledge Ingestion

Data is processed at three levels: raw external data acquisition, basic cleaning and fusion, and deep business‑driven mining. This results in a high‑value intelligence system divided into efficiency, incremental, and decision‑making modules, supporting over 10 business lines and dozens of centers.

3. Knowledge Organization

The knowledge is categorized into five domains: entity basics (e.g., property listings), operational standards, company policies, policy interpretations, and cross‑entity relationships, forming a graph with over 60 entity types and a scale approaching 500 billion records.

4. Reasoning Graph Progress

To handle complex queries such as detailed school district eligibility, a reasoning graph is built on top of the entity graph, enabling multi‑level rule inference for scenarios like tax calculations, achieving near‑100% accuracy across dozens of cities.

5. Application in Intelligent Training

The knowledge graph is deployed to a training platform that provides standardized Q&A, evaluation, and feedback loops, improving agent skill levels and reducing training costs. The platform evolved through four stages: productization, platformization, scenario‑based VR training, and modularization for broader business reuse.

6. Knowledge Operation Loop

An operation mechanism monitors data quality through empty‑value detection, deviation detection, user guidance, proactive probing, and sampling annotation, achieving an estimated data accuracy ceiling of about 98%.

7. Future Outlook

Beike plans to solidify its role as the industry knowledge hub, extending standards beyond internal use to the broader real‑estate sector, and to deepen the training platform into a comprehensive growth pathway for agents.

Thank you for listening.