Meituan Recommendation System Architecture and Optimization Practices

Recommendation systems have long existed, but their widespread adoption in internet companies is a recent phenomenon. With the explosion of online information, users face severe information overload, making personalized recommendation a crucial solution.

Meituan's recommendation platform is divided into four layers: data layer (log cleaning and storage), candidate generation layer (triggering based on user behavior, real‑time actions, and location), fusion & filtering layer (combining multiple candidate sets and applying business rules), and ranking layer (machine‑learning models re‑order candidates). Both the trigger and ranking layers are decoupled to support independent A/B testing.

Data is the foundation for algorithms and models. Meituan leverages massive user behavior logs, distinguishing between active actions (browsing, ordering), negative feedback, user profiles, and UGC tags to feed offline calculations and online predictions.

Trigger strategies include:

Collaborative filtering (user‑based, item‑based) with data cleaning, time‑window selection, and similarity measures such as log‑likelihood ratio.

Location‑based triggers that exploit real‑time, work, and home locations to weight regional hot deals.

Query‑based triggers that assign weights to historical queries and associated deals.

Graph‑based triggers using bipartite user‑deal graphs and algorithms like SimRank.

Real‑time behavior triggers (real‑time browse, favorite).

Supplementary strategies for cold‑start users (hot‑selling items, high‑rating items, city‑specific items).

These sub‑strategies are fused using weighted, hierarchical, modulation, and filtering methods; Meituan currently combines modulation and hierarchical fusion.

Candidate re‑ranking relies on machine‑learning models:

Non‑linear tree ensembles (Additive Groves) that aggregate multiple groves via bagging.

Linear models, primarily Logistic Regression, updated online with Google's FTRL algorithm, processing feature vectors stored in HBase and streamed via Storm.

Data handling includes sampling to address click‑through rate imbalance, careful definition of negative samples, and noise removal (e.g., fraud detection).

Features used in ranking are categorized as deal‑level (price, discount, sales, rating), user‑level (level, demographics, client type), cross features (user‑deal interactions), and distance features (geolocation relative to POIs). Non‑linear models accept raw features, while linear models require bucketization and normalization.

In summary, Meituan improves recommendation performance by fusing diverse candidate generation strategies and applying sophisticated ranking models, demonstrating the synergy between data engineering and algorithmic innovation.