Meituan-Dianping Tech Salon: Full‑Chain Application of Food‑Delivery Big Data – User Profiling, Marketing Strategies, and Predictive Modeling

This article is compiled from the 9th session of the Meituan-Dianping Technology Salon titled “Food‑Delivery Big Data – Full‑Chain Application Reveal”.

Speaker: Li Tao, Senior Technical Expert at Meituan-Dianping, responsible for user profiling and marketing strategy research. He previously worked at Ricoh Beijing Research Institute and Teradata, where he designed the first‑generation face detection/focus system for Ricoh cameras and developed large‑scale data‑mining algorithms on Teradata Aster.

Presentation Overview: The talk is divided into four parts: (1) Marketing requirements of the food‑delivery business; (2) Meituan user profiling; (3) Case studies illustrating common problems and practical experiences; (4) Personal reflections and conclusions.

The user journey is split into two stages: acquisition and expansion. In the acquisition stage, personalized ads and promotions attract first‑time orders. After the first order, the focus shifts to encouraging second and third purchases, often through cross‑selling (e.g., recommending night‑snack or afternoon‑tea items).

To determine a user’s stage, a data‑driven RFM matrix is used, segmenting users into: growth users (new), mature users (frequent), churn‑risk users (inactive for a period), and “dead” users (long‑term inactivity). Each segment has specific operational strategies: increase order value for growth users, boost value for mature users, retain churn‑risk users via churn‑prediction, and selectively recall dead users.

Beyond stage identification, a comprehensive user profile includes demographic information (age, gender, occupation, household assets), preferences (spiciness, sweetness, short‑term vs. long‑term tastes), consumption price, ordering time and scenario, and behavior derived from group‑buy and review data (e.g., sentiment extraction, merchant and dish characteristics).

These profiling data are stored in tables and exposed as services for downstream business systems. The talk then shares practical experiences in three key operational areas:

Acquisition (拉新): Identify potential food‑delivery users by analyzing demographic attributes, shared preferences within offices, and consumption potential inferred from ordering locations and habits.

Retention (留存): Build churn‑prediction models by extracting churn‑related features and training a probability‑regression model (predicting the likelihood of a user not ordering in the next few days) and a survival model (predicting the time until the next order). Both models show comparable performance, with the survival model providing an estimated interval.

Experience Expansion (拓展): Understand ordering scenarios from three dimensions: time, location, and order content. Scenarios include weekday afternoon tea, weekend family meals, overtime meals, travel, internet‑café orders, or late‑night snacks. Scene construction involves basic feature extraction from order attributes, followed by qualitative interviews, user segmentation, multivariate analysis, and data‑mining to discover and define scenes. Users are then classified into scenes using rule‑based or predictive algorithms, often supported by cloud‑computing services.

The food‑delivery industry’s marketing characteristics differ from other sectors mainly in two aspects: (1) a strong focus on the user lifecycle due to high transaction frequency, and (2) scenario‑driven demand. Challenges include processing massive unstructured data (orders, dish names), limited user awareness of food‑delivery services, and short decision‑making windows with scarce pre‑order signals, requiring inference from behavioral data.

Meituan-Dianping Technology Salon is organized by the Meituan‑Dianping technical team, held monthly, inviting experts from Meituan‑Dianping and other internet companies to share frontline practice across major technology fields.