Data‑Driven Precise Marketing: Architecture and Case Studies from Meituan Dianping

Precise marketing has long been a powerful tool for acquiring users and boosting conversion in segmented internet markets. With the explosion of mobile internet, data volumes have grown exponentially, making data‑driven precise marketing a major challenge and a key research direction in big‑data applications. This article shares the data architecture and technical implementations used by Meituan Dianping's data application team.

Overall Framework O2O marketing consists of channels (online, offline) and business themes (traffic, transactions). Various forms such as precise user campaigns, DSP targeting, channel ranking, and anti‑fraud rely heavily on data analysis. The article focuses on in‑site precise user marketing, which follows six stages: goal definition, audience selection, campaign design, configuration & launch, online optimization, and effect monitoring.

The team identified three key problems: (1) pre‑campaign goal setting and audience selection lacked stable analytical frameworks; (2) real‑time effect monitoring and strategy output were missing; (3) post‑campaign evaluation suffered from inconsistent metrics.

System Architecture A layered pyramid architecture was designed to meet these needs, featuring a data warehouse and model layer at the bottom, a service layer in the middle, and data products/applications on top. The warehouse stores three themes—user profiles, operations & marketing, and traffic—providing essential data for campaigns.

Profile data combines self‑built tags with those from search, advertising, and risk‑control teams, covering over 180 tags across five categories (basic info, device, consumption, browsing, and demographic groups). The team evolved from simple statistical models (e.g., RFM) to machine‑learning‑based user preference mining.

For budget control, a company‑wide budget serial‑number system was built, linking financial approvals to discount back‑ends, enabling fine‑grained monitoring of each order. Metrics such as new‑user cost, retention, and purchase frequency were derived from the modeled data.

Traffic metrics (PV, UV, session, path‑tree conversion) were implemented using Elasticsearch after evaluating Kylin and Druid, due to ES's lower storage overhead and simpler operations.

Data services use a high‑performance RPC framework; profile services require millisecond latency and thus use Redis, while analytical services favor ES for its aggregation capabilities.

Data Products & Tools

Audience Analysis Platform (Hoek): enables rapid creation of audience segments using profile tags, which can be fed to push or promotion tools.

Smart Coupon Engine (Cord): a configurable, service‑oriented engine that lets operators define target audiences and strategies without code, supporting A/B testing and dynamic recommendation.

CloudMap/StarMap: ES‑based query engine providing real‑time multi‑dimensional metric dashboards for campaign performance.

Through more than 20 thematic analyses—including dynamic budget allocation, discount gradient optimization, and user‑value scoring—the team improved budget utilization by 30% and better assessed user acquisition value.

Case Studies

Potential User Mining for Takeaway With nearly 1 billion active users across ~20 business lines, the team explored look‑alike user discovery using association rules, clustering, and classification models. Results showed significant improvements over baseline methods, and future work will incorporate social‑graph propagation via Spark GraphX.

WeChat Red‑Packet Precise Coupon Engine Cord treats coupon distribution as a simplified recommendation problem. It consists of a traffic‑splitting module (gray release & A/B testing), a recall module (fetching user and coupon assets), a filtering module (matching), and a ranking module (applying business rules or mined strategies). The system is fully service‑oriented and configurable, allowing external campaign systems to enable or disable Cord dynamically. Scoring combines user preference and coupon‑sensitivity tags to support both GMV‑driven and user‑acquisition‑driven recommendation strategies.

Summary Precise marketing is a recognized big‑data scenario; as mobile internet and O2O evolve, new challenges arise. The team distilled design principles: (1) Build accurate, easy‑to‑use bottom‑layer data models and unified metrics from business needs; (2) Apply layered SOA architecture to decouple services and select appropriate technologies per scenario.

Future work will focus on rapid model development and expanding the architecture to support more marketing scenarios, further unlocking data value.