User Profiling and Machine Learning Practices for Food Delivery O2O Platforms

Meituan Delivery has rapidly expanded over three years from a single‑category food‑delivery service to include meals, night‑snacks, flowers, supermarkets and more, reaching diverse user groups such as students, white‑collar workers, communities and business travelers. Building accurate user portraits is essential to match supply with demand, capturing each group’s consumption habits and preferences.

Compared with traditional e‑commerce, food‑delivery O2O differs in four key ways: it is a new, fast‑growing service; it has high transaction frequency with low order value and short decision time; it is driven by specific consumption scenarios (time, location, user type); and user location is relatively fixed, making geographic context a strong indicator of intent.

The product operation can be divided into two stages. In the acquisition stage, the goal is to attract high‑potential users through personalized marketing and convert them to first‑time buyers, thereby increasing LTV. In the expansion stage, the focus shifts to raising average order value and purchase frequency via cross‑selling, upselling, and retention tactics such as coupons and loyalty incentives.

System Architecture Data sources (basic logs, merchant data, order data) are processed into topic tables, then stored in KV stores for online services. Two monitoring layers track data‑processing health and service performance. Downstream services include advertising, ranking and operation systems.

New‑Customer Operation Key questions are: where are the new customers, what are their preferences, and what is their purchasing power. Sparse behavior requires inference from demographic attributes, location, surrounding crowd, and supply availability. Feature engineering and smoothing techniques are applied to build a high‑potential‑customer prediction model, achieving conversion rates several times higher than average.

Churn Prediction Retaining users after acquisition is critical. Two modeling approaches are used: (1) probability regression (e.g., logistic regression, decision trees) to predict the likelihood of an order within a future time window, and (2) survival analysis using the Cox‑PH model to estimate churn risk. The probability‑regression diagram and Cox model example are shown below.

The Cox model’s βTx coefficient correlates with the time interval between orders, as illustrated in the box‑plot diagram.

Both models achieve comparable performance in predicting whether a user will place an order within a given horizon, and Meituan’s churn‑warning system has significantly reduced retention costs.

Scenario‑Based Operation Understanding the order scenario (time, location, order type) enables targeted cross‑selling and upselling. The workflow includes scenario identification, address analysis (extracting core address and type via NLP and map data), and merging orders to detect group orders. This information feeds into scenario prediction and personalized marketing.

Summary Food‑delivery marketing is distinguished by high frequency, clear lifecycle stages, and strong scenario influence. Effective profiling combines basic attributes, preferences, purchasing power, and churn risk to drive precise recommendations and pricing strategies. Challenges include diverse unstructured data (e.g., addresses, dish names), sparse user behavior, and low order value, requiring advanced NLP, data fusion, user segmentation, look‑alike modeling, and transfer learning. Overcoming these challenges can advance big‑data applications across e‑commerce domains.