Evolution and Architecture of Beike's Intelligent Real Estate Recommendation Platform

Xu Yansong, a senior algorithm expert at Beike, presented at ArchSummit 2018 a comprehensive overview of the company's intelligent recommendation platform, explaining how architectural upgrades improve iteration speed, how diverse algorithms boost recommendation quality, and sharing practical lessons learned.

The platform serves both home‑buyers during the house‑search phase and agents during lead‑management, aiming to increase browsing efficiency for users and provide targeted property suggestions for agents.

Version 1.0 relied on simple content‑based and rule‑based strategies (e.g., new‑listing, price‑drop rules) using only basic property data, which lacked personalization and suffered from tightly coupled code and limited data sources.

Version 2.0 introduced personalization through collaborative‑filtering (using abstracted house tags to mitigate sparsity) and user‑profile‑based recommendations, added a recommendation metric system, and began restructuring the offline‑mining pipeline into layered data, algorithm, and evaluation components.

Version 3.0 reorganized the online service into three layers—application, data, and compute—exposed a unified API, made strategy configuration JSON‑based, and enabled rapid business onboarding, flexible rule‑based strategy selection, and fast iteration.

Stability improvements include configurable fallback strategies, an intelligent circuit‑breaker for unstable data sources, independent scaling of the three layers, and achieving five‑nine SLA reliability, complemented by comprehensive monitoring of strategy health and service metrics.

A closed‑loop data ecosystem was built: recommendation requests trigger AB‑group routing, candidate generation, scoring, exposure tagging, logging, real‑time event collection via Kafka, Spark processing, and feedback storage in Redis, which together raise CTR by over 10%.

Future work focuses on deeper algorithmic research (e.g., WDL), automated online learning to replace manual AB tests, supporting more complex models, and extending the platform beyond Beike to external advertising scenarios.

The speaker concluded with reflections on the importance of solid architecture, data‑driven decision making, and continuous experimentation.