Beike Commercialization Strategy Algorithm Platform: Architecture, AI Techniques, and System Evolution

The talk begins with an overview of Beike's commercial business model, where real‑estate agents act as advertisers and earn "opportunity" (商机) and "commission" (委托) through a bidding system based on ECPM, pCVR, and CPA settlement.

Four main parts are covered: an introduction to Beike's commercialization, the architecture evolution (2018‑2020), the design of the new strategy algorithm platform, and a summary with outlook.

Architecture evolution highlights three stages: 2018’s coarse‑grained system using CPT orders, Hive tables, and Redis; 2019’s addition of effect advertising and Python‑based CVR/ECPM models; and 2020’s rapid product growth exposing performance, stability, data redundancy, and onboarding bottlenecks.

To address these issues, the platform was rebuilt into a three‑layer design: an online service layer for low‑latency HTTP APIs and A/B testing, a near‑real‑time layer for feature updates, vector retrieval, and streaming data, and an offline layer for large‑scale batch processing, feature engineering, and model training (LR, LightGBM, Wide&Deep, DeepFM).

Key technical solutions include migrating performance‑critical services from Python to Java, introducing service‑governance (circuit breaking, rate limiting), modularizing micro‑services via Spring Cloud, using Eureka/Feign/Ribbon for RPC, and employing Prometheus/Grafana for monitoring.

Vector search transitioned from Faiss (single‑node) to Milvus for distributed similarity retrieval, with embeddings generated via node2vec on user‑item interaction graphs.

Model serving utilizes PMML for simple models and an internal RunOnce platform (TensorFlow/PyTorch) for complex models, integrated with Spark and Kafka pipelines.

After reconstruction, system stability improved from 97% to 99.99%, resource usage dropped (fewer instances), and operational efficiency increased, enabling rapid configuration‑driven feature rollout and real‑time KPI dashboards.

The outlook envisions expanding the platform into a full AI middle‑platform supporting additional services such as image retrieval and property valuation, broader language support, further RPC performance optimization, and continuous model innovation.