Meituan Delivery Big Data: Full‑Chain Application Insights

This article is compiled from Meituan‑Dianping Technical Salon Session 9 – “Delivery Big Data – Full‑Chain Application Reveal”.

Speaker: Wang Xingxing, Senior Technical Expert at Meituan‑Dianping, responsible for delivery advertising technology. Former senior researcher at Sogou Advertising, with experience in feature frameworks and training systems, and award‑winning data‑mining competitions.

Key Characteristics of Delivery Compared to Group Buying:

1. Mobile‑centric : Since 2011 the mobile strategy grew, and by 2014 mobile accounted for over 75% of delivery usage, tightly linked to phone‑based ordering and delivery communication.

2. Local‑centric : The dominant category is food, with about 65% of orders within a 1‑km radius, imposing strong constraints on technical choices.

3. Scenario‑centric : Daily order volume shows two clear peaks corresponding to lunch and dinner periods.

From a city‑level perspective, order forecasting faces massive manual verification across hundreds of cities, making it difficult to identify the root causes of fluctuations and quantify their impact.

Modeling Objectives: (1) Automatic monitoring and alerting; (2) Detect anomalies and pinpoint their causes; (3) Quantify the impact of each cause on order volume.

The solution uses a two‑stage modeling pipeline: first, a historical autoregressive model predicts baseline orders; second, a “multiplier” model (using the ratio of actual to predicted as label) refines the forecast, leveraging boosting for higher accuracy and providing better interpretability.

List Optimization Architecture: The pipeline consists of three layers – (1) Log collection (source of data and strategy); (2) Offline processing (analysis, model training); (3) Online service (API requests from front‑end, including AB‑test framework, trigger and ranking modules, and real‑time data processing).

During model training, selecting negative samples is a challenge because only positive feedback is explicit. Strategies explored include random selection, choosing popular items with no user interaction, and sampling from candidate sets of highly active users. The “Skip‑above” method performed well for the delivery list scenario.

Feature System: Features are organized into three dimensions – scene, user, and product – and their pairwise intersections, forming seven quadrants. In delivery, scene features are relatively weak; the main features fall into three categories: merchant‑level, user‑level, and merchant‑user interaction.

Tree‑based models excel at handling dense, enumerated features (e.g., ratings, distances) that linear models struggle with, which explains their popularity at Meituan.

Real‑Time Solutions: Three approaches are used: (1) Model‑driven inference for non‑statistical features; (2) Frequent feature updates for statistical features (e.g., dynamic merchant promotions); (3) Hybrid of both.

Delivery introduces unique challenges such as strong logistics constraints. For example, when a merchant’s delivery time lengthens, the ranking results must reflect this change.

Cold‑start problems for new merchants and new users are also difficult. Proposed solutions include: (1) Fixed‑position slots with predefined rules for new merchants; (2) “E&E” (exposure & evaluation) methods to differentiate performance over time; (3) Providing a traffic entry (e.g., paid advertising) to acquire initial exposure.

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