Engineering Architecture Practices for an AI‑Powered Recommendation Platform at Beike

1 Introduction

Beike senior engineer Yuan Bin shares the engineering practice of the company's AI‑driven intelligent recommendation platform, covering four parts: platform overview, evolution roadmap, architecture design, and future outlook.

2 Platform Overview

The platform serves both C‑end users (home‑buyers) and B‑end agents, supporting more than 300 recommendation scenarios across second‑hand houses, new homes, rentals, and content feeds. For C‑end users it improves house‑search efficiency during the "find house" stage; for agents it accelerates the "maintenance and marketing" stage of the sales workflow.

3 Evolution Roadmap

The architecture has progressed through four versions:

V1.0 – rapid construction with separate online and offline modules, rule‑based and popularity‑based strategies.

V2.0 – behavior collection, offline reconstruction, modularized offline pipeline (data, feature, algorithm, evaluation, output) and introduction of CTR/ROI metrics ( CTR = click_count / impression_count , ROI = click_count / module_impression_count ).

V3.0 – three‑layer service (application, computation, data), strategy configuration, real‑time streaming via Spark Streaming, and a generic real‑time framework.

V4.0 – addition of a model layer for complex deep‑learning predictions, expanded strategy stages, and enhanced explainability.

4 Architecture Design

The design follows three principles: stage separation, appropriate reservation, and configuration programming. Strategies are split into blacklist (filter), recall, fusion, ranking, and reason stages, enabling clear responsibilities and easy composition.

Recall strategy configuration includes extra‑information, query building, execution, and result parsing, with examples of ID‑based context features for home‑detail and homepage recommendations.

Configuration programming reduces code from hundreds of lines to a few dozen configuration lines, improving development efficiency and reducing risk.

5 Future Plans

Beike will continue to focus on cost reduction and efficiency, pursuing search‑recommendation architecture fusion and establishing a common core‑strategy framework that decouples strategy from business, allowing rapid reuse across scenarios.

By the end of the presentation the platform supports 319 scenarios, over 1,000 distinct strategies, handles 250 million daily requests, and maintains a 99.999 % SLA.