How Meituan’s Data Architecture Powers Precise Mobile Marketing

Precise marketing has long been a key tool for quickly acquiring users and improving conversion in segmented markets. With the explosion of mobile internet and exponential data growth, leveraging data‑driven precise marketing in mobile and big‑data scenarios is a major challenge and research direction.

Overall Framework

Before introducing the data system, it is helpful to understand the basic components of O2O marketing: it consists of two dimensions—marketing channels (on‑site, off‑site) and marketing themes (traffic, transaction). Various forms such as precise user marketing activities, DSP targeting, channel ranking, and anti‑fraud all benefit from data analysis. The article focuses on on‑site precise user marketing activities.

An on‑site user operation activity typically follows six stages: defining goals, selecting target users, designing the activity plan, configuring and launching, online precise marketing with dynamic optimization, and finally monitoring and evaluating results.

System Architecture

Based on the business scenarios and requirements, a layered pyramid architecture was designed to build a marketing data system and services that meet business needs while offering strong platform extensibility.

The lowest layer is the data warehouse and model layer, which is divided into three themes: user profiling, operation & marketing, and traffic—each essential for operational activities.

For profiling, a hybrid approach combines self‑built tags with tags from search, advertising, and risk‑control teams, forming a unified wide‑table. Over 180 tags across five categories (basic info, device info, consumption & browsing, demographic groups, etc.) are maintained. The implementation evolved from simple statistical and RFM models to machine‑learning‑based user‑interest mining.

In collaboration with finance and payment systems, a budget serial‑number system was built. Operators request a budget serial number from finance, link it to the activity configuration, and each order triggers a data point, enabling fine‑grained budget monitoring. After modeling users and products, a marketing transaction evaluation metric suite was established (e.g., new‑user cost, new/old user distribution, 7‑day and 30‑day purchase retention).

Traffic metrics such as page clicks, conversion funnels, and channel sources are also captured. Core models for PV, UV, session, and path‑tree conversion were built in the data warehouse to satisfy operational needs.

On top of the warehouse, a data service layer was created using a high‑performance RPC framework. Different storage and query engines were chosen per use case: Redis for real‑time profiling services requiring millisecond latency, and Elasticsearch for analytical workloads because it offers lower storage overhead than Kylin and simpler operations.

Audience Analysis Platform (Hoek): Users create audience packages by combining profiling tags, which can be fed to various marketing tools such as push notifications and promotions.

Smart Coupon Engine (Cord): Operators configure targeted audiences and strategies via a backend and Hoek without any development effort.

CloudMap/StarMap: A tool platform built on Elasticsearch that provides multi‑dimensional real‑time metric queries for activity effect analysis.

Beyond systematic construction, more than 20 specialized analyses were conducted, producing models like dynamic budget allocation, discount gradient optimization, user value scoring, and free‑entertainment recommendation, improving budget utilization by 30% and better evaluating user acquisition value.

Case Sharing

Potential User Mining for Delivery

Precise marketing aims to discover potential customers among nearly 100 million active users. While platforms like Facebook and Tencent provide Look‑alike functions, Meituan’s approach combines profiling, association rules, clustering, and classification models. The following figure compares the effectiveness of these methods.

Future work includes leveraging user‑friend relationships and Spark GraphX to implement label‑propagation algorithms for deeper similarity mining.

WeChat Red Packet Precise Coupon Engine

The Cord engine addresses the problem of selecting the most effective coupon for a user in WeChat groups. It follows a recommendation‑style pipeline: a traffic‑splitting module for gray releases and A/B testing, a recall module that fetches user and coupon information from profiling and configuration services, a filtering module that matches them, and a recommendation module that ranks results based on business rules or mined strategies (e.g., GMV‑oriented or user‑acquisition‑oriented). The system is fully service‑oriented and configurable, allowing external activity systems to enable or disable Cord via feature flags.

Conclusion

Build accurate, easy‑to‑use bottom‑level data models and a unified metric system driven by business requirements.

Apply layered SOA principles to decouple data services from business logic, selecting appropriate technology components for each scenario.

Looking ahead, the focus will be on accelerating model development, expanding architectural support for more scenarios, and unlocking greater marketing value from data.