Meituan Search Ranking: Multi‑Business Sorting Architecture and Optimization Practices

Meituan operates a diverse set of business categories, and its search service must return heterogeneous results that may include store POIs, delivery POIs, selected products, and more, requiring a ranking architecture that can handle many business scenarios.

The ranking framework is divided into four layers: an entry presentation layer, a heterogeneous merging layer that combines items of four types (goods, merchants, ads, cards), a multi‑stage sorting layer (L1 coarse recall, L2 fine ranking, L3 strategy models) with separate homogeneous sorting for goods and merchants, and a foundational retrieval layer that provides business‑specific indexes and recall services.

To support rapid algorithm iteration, Meituan built several platforms, including the Lego orchestration platform for constructing ranking services, and the Poker one‑stop training and experimentation platform that covers data acquisition, feature management, model training, evaluation, and A/B testing.

Multi‑business modeling is addressed through a quota model (MQM) that determines how many items from each recall path advance to the next stage, using a cross of recall method and business type as targets, enhanced with Transformer‑based sequence modeling and business priors.

The fine‑ranking model (MBN) evolved through five versions, progressively splitting business‑specific sub‑networks, introducing learnable sub‑network weights, applying MMoE for adaptive feature sharing, adopting CGC for clearer expert separation, and finally incorporating a probabilistic graph with prior and posterior networks to better capture multi‑task objectives.

Aggregation modeling predicts both the position and size of result blocks (e.g., fresh produce, Meituan Select) using a joint multi‑objective approach that handles heterogeneous item‑level and block‑level behavior sequences with separate attention mechanisms.

The presentation concludes with a summary of the work on business understanding, data capabilities, architectural layering, and system engineering, followed by a Q&A session, recruitment invitation, and links to additional resources.