Knowledge Graph and Distributed Graph Database Practices at Beike Zhaofang

Recently, Beike Zhaofang held a Knowledge Graph Technology Conference in Beijing, where four senior engineers presented the practical experience and results of applying graph technology in the company's products.

Senior engineer Wang Xuezhi explained the deployment of relationship graphs in Beike's risk‑control system, describing the ACN broker cooperation network, the unique risk‑control scenarios, and why graph‑based risk modeling is suitable for detecting organized fraud.

The relationship‑graph architecture consists of four layers—basic data, knowledge construction, knowledge mining, and business application—using Spark GraphX for graph computation and Janus Graph for querying. Applications include admission control, risk quantification, quality management, risk discovery, and case tracing.

Future work will focus on enhancing core capabilities such as knowledge reasoning, knowledge fusion, high‑density sub‑graph mining, and graph embedding, as well as expanding business use cases like automated violation tracing and broker credit‑based recommendations.

Algorithm engineer Zhou Yuchi presented the design and implementation of Beike's overall relationship‑graph system, covering the motivations, the four‑layer architecture (basic graph, sub‑graph, graph capabilities, graph applications), and the quantification of relationship strength based on weight, frequency, and recency.

Key graph capabilities highlighted include influence scoring, graph embedding (Node2vec for homogeneous graphs, Metapath2vec for heterogeneous graphs), similarity calculation, and relationship prediction, which support services such as the internal "Fangke Tong" platform and intelligent recommendation engines.

Search platform lead Gao Pan discussed the need for a distributed graph database to store the massive 48‑billion triple dataset, comparing DGraph and JanusGraph and ultimately selecting DGraph for its performance, achieving sub‑50 ms query latency and over 15 000 QPS.

Gao also outlined future optimization plans for the graph database, its integration as a core search‑engine component, and the goal of providing a unified graph‑database platform for all knowledge‑graph and risk‑graph needs across the company.

Industry‑graph leader Sun Baqun described how internal and external data are fused to build an industry knowledge graph that supports strategic planning, intelligent Q&A, knowledge reasoning, and market intelligence, ultimately driving business growth and improving user experience.