How Meituan Supercharges Local Services with Advanced Recommendation and Ranking

Meituan Recommendation Products

Shen Guoyang introduces Meituan's major recommendation offerings:

Guess You Like : The flagship product that aims to present the most relevant group‑buy services when users open the app, increasing transaction share from 0.7% to 7‑8%.

Home Channel Recommendations : Fixed and personalized channels that contribute over 40% of Meituan's transaction volume.

Today's Personalized Push : Pushes the most interesting services to users before they open the app, driving clicks and orders.

Category List Personalized Ranking : Intelligent ordering of items within category channels, offering higher personalization than search but slightly lower than the "Guess You Like" slot.

Goals of Meituan's Recommendation System

The system aims to help users quickly find high‑quality, low‑price services, measured primarily by post‑recommendation order conversion. It also seeks diversity across multiple categories to reinforce Meituan's brand as a comprehensive "food, drink, play, and travel" platform.

Overall Framework

The top layer exposes service interfaces for different placement slots, each with its own API. An AB‑test configuration module routes traffic by UUID, city, and other dimensions, allowing real‑time, no‑restart adjustments.

Below the AB‑test layer are candidate generation, ranking, and business‑logic modules. Candidate and ranking strategies can be configured per placement, while business rules have both shared and placement‑specific logic.

During request handling, logs capture contextual information and user/item features, which are streamed via Flume to HDFS. Using Hadoop, Hive, Spark, and Meituan's own machine‑learning libraries, raw logs are processed into profiles, similarity matrices, geographic‑item relations, and conversion‑rate prediction models.

These data and models feed the candidate generation, ranking, and business‑processing modules.

Key Algorithms and Characteristics

Meituan heavily employs traditional collaborative‑filtering methods (user‑based and item‑based) with a time‑decay factor that emphasizes recent behavior. Log‑likelihood‑ratio (LLR) similarity outperforms cosine similarity, and user‑based algorithms work better for the home‑page "Guess You Like" slot due to higher novelty.

1. Cold‑Start Challenge

Because many users have sparse histories, Meituan implements cold‑start strategies, such as the "Local Hot Items" approach that promotes top items within a user's real‑time business district (average coverage of a few dozen square kilometers).

2. Mobile‑Centric Usage

Over 90% of transactions occur on mobile devices, leading to usage patterns like checking nearby restaurants after shopping or selecting a hotel after a movie.

3. Short Holding Time

For high‑frequency categories (food, movies), 40% of users consume within an hour, while low‑frequency categories (photography, hairdressing) have longer intervals.

4. Short Holding Distance

For most cities and categories, 80% of orders are placed within 2 km of the consumption location.

Major Ranking Work

1. Modeling

Meituan uses Additive Groves (a tree‑based, non‑linear model) and explores online‑learning models like FTRL. The approach is point‑wise, treating each user‑item impression as a sample with click or order as the label, and extracts user, item, and context features.

2. Sampling and Label Processing

Orders are the primary conversion signal, but to mitigate sparsity, clicks are also used with different weighting schemes (e.g., up‑sampling order samples 30× or assigning higher label values to orders).

3. Position Bias Removal

Position bias is mitigated by normalizing historical CTR/CVR per position and by setting position features to zero during training and inference.

4. Feature Engineering

Key features include context (time, location, weather), item attributes (price, sales, rating), and user demographics (age, gender, preferences).

5. Interleaving Evaluation

Beyond AB‑tests, Meituan employs interleaving (Balanced method) to compare two ranking strategies with lower traffic and faster feedback, measuring wins, ties, and deltas.

6. Online Learning Attempts

Recognizing that user behavior shifts with seasons, weather, trends, and events, Meituan is experimenting with online learning to capture these dynamics, though stability remains a challenge.