Applying Knowledge Graph Technology to Real Estate Search: Product Overview and Technical Architecture

Introduction – "Kelu Fang" is an innovative product incubated during the second Beike Find House Hackathon, applying knowledge‑graph technology to a map‑based real‑estate search experience. The article outlines the product’s background, design, technical implementation, and future outlook.

Background – Analysis of Beike’s data revealed that many users are commuters who search for housing around their workplace, caring about commute time, transportation mode, and surrounding amenities such as supermarkets, restaurants, and schools. To address these needs, two enhancements were proposed: a commute‑filter module on the main page and a scalable local view showing nearby facilities, both powered by a knowledge graph.

Product Effect – The web‑based UI adds time‑range and transportation‑mode selectors, allowing users to filter listings by commute criteria directly on the homepage. After selecting a listing, a graph‑based view displays surrounding facilities with detailed tags (e.g., "Three‑A Hospital"). The map also supports direct location search, commute‑time calculation for four transport modes, and visualizes routes and nearby amenities.

Technical Solution

The system consists of four layers: data, mining, computation, and front‑end UI.

4.1 Data Layer – Combines structured data from the property dictionary and logs with semi‑structured/unstructured data such as Q&A and articles.

4.2 Mining Layer – Utilizes a real‑estate knowledge graph and an expanded NLU lexicon. User search and click logs build user profiles for personalized recommendation. The knowledge graph stores two types of triples:

Attribute triples <entity attribute value>

Relation triples <entity relation entity>

It currently covers 40 entity types, ~190 relations, 1.6 billion entities, and 39 billion triples, making it well‑suited for AI‑driven graph analytics.

The NLU lexicon extends entity names with over 70 k community terms and more than 1.53 million aliases (e.g., "Beijing University Third Hospital" → "北医三院"). This improves recall for user queries.

High‑quality POI data (≈15 million nationwide, 300 k in Beijing) are pre‑computed for four commute modes within a 20 km radius. Each record follows a CVT (Compound Value Type) model containing time, distance, and mode, and is stored in Neo4j for fast retrieval.

4.3 Computation Service Layer – Applies NLU to interpret user intent (location intent and commute intent) using lexicon matching and a BiLSTM‑CRF model. Retrieval ranking disambiguates place names by estimating the probability of each place type using Bayes’ rule:

t' = argmax(p(t|c)) = argmax( p(c|t)p(t))

Click‑log statistics provide the priors p(t) and likelihoods p(c|t), enabling re‑ranking of search results.

Future Outlook – Planned improvements include a more refined local view with hierarchical facility categories, richer knowledge‑graph inference beyond structured data, expanded tagging of listings using advanced extraction algorithms, and personalized ranking based on detailed user profiles.

Conclusion – "Kelu Fang" demonstrates a successful application of knowledge‑graph technology in real‑estate search, offering enhanced commute filtering and contextual facility visualization. While functional, further user‑centered research and technical refinements are needed to boost the overall experience.