Multi‑Business Ranking Modeling and Optimization in Meituan Search

Introduction

Meituan's mission is to help people eat and live better, and its app hosts many services such as food delivery, in‑store dining, grocery, hotels, travel, and entertainment. The homepage search is the main traffic entry, often returning mixed‑business results (e.g., restaurants, movies, hotels) for ambiguous queries like "Wangjing". This creates challenges: differing feature sensitivities across businesses, data imbalance between high‑frequency and low‑frequency services, and distinct business objectives.

Ranking Process Overview

The overall pipeline consists of data, recall, ranking, and presentation layers. The ranking layer is divided into coarse ranking, multi‑path fusion, fine ranking, re‑ranking, and heterogeneous ranking. The article focuses on the multi‑path fusion and fine‑ranking stages.

Multi‑Business Modeling Practice

Multi‑Business Quota Model (Multi‑Path Fusion Layer)

To balance the proportion of candidates from different businesses, Meituan designed a Multi‑Business Quota Model (MQM). MQM‑V1 uses a multi‑objective formulation (click + order) with query, context, cross, and user features, outputting joint probabilities for each business. It improved overall click‑through rate by +0.53%.

MQM‑V2 adds a two‑dimensional objective (recall × business), introduces a Transformer‑based behavior‑sequence module, and incorporates an expert layer for cold‑start handling. This version raised tourism purchase rate by +2% and in‑store dining by +0.57%.

Multi‑Business Ranking Model (Fine‑Ranking Layer)

The fine‑ranking stage evolved from a standard Embedding&MLP DNN to specialized multi‑task architectures named Multi‑Business Network (MBN).

MBN‑V1

Introduced independent sub‑networks for hotel and travel, combined with a shared main network; hard‑split (one‑hot) and soft‑split (quota‑based) weighting were compared, with soft‑split yielding better results (+0.17% overall CTR).

MBN‑V2

Added a food sub‑network and decoupled the quota model by embedding a weight‑generation sub‑network that consumes query and context features. This version improved overall CTR by +0.1%.

MBN‑V3

Adopted a Multi‑gate Mixture‑of‑Experts (MMoE) structure for the new “to‑review” sub‑network, added a business‑specific cross‑entropy loss, and achieved notable gains for food (+0.36%), review (+1.07%), hotel (+0.27%), and travel (+0.35%).

MBN‑V4

Replaced MMoE with Progressive Layered Extraction (PLE/CGC) to give each task dedicated experts while sharing others, leading to more stable expert weights and further improvements (overall CTR +0.1%, food +0.53%).

Summary and Outlook

The multi‑business ranking system at Meituan has progressed through extensive engineering, algorithmic, and product innovations, especially in the multi‑path fusion and fine‑ranking layers. Future work includes better utilization of business‑specific features, handling data imbalance via transfer or meta‑learning, and multi‑objective optimization to align user experience with diverse business goals.

References

[1] Product‑based Neural Networks for User Response Prediction [2] DeepFM: A Factorization‑Machine based Neural Network for CTR Prediction [3] Deep & Cross Network for Ad Click Predictions [4] AutoInt: Automatic Feature Interaction Learning via Self‑Attentive Neural Networks [5] FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for CTR Prediction [6] Transformer in Meituan Search Ranking Practice [7] Modeling Task Relationships in Multi‑task Learning with Multi‑gate Mixture‑of‑Experts [8] Progressive Layered Extraction (PLE): A Novel Multi‑Task Learning Model for Personalized Recommendations

Author Bio

Pai Hao, Xiao Yao, Xiao Jiang, Jia Qi, Chen Sheng, Yun Sen, Yong Chao, Li Qian, etc., all from Meituan's Search & NLP team.