Multi‑Business Product Ranking in Meituan Search: Challenges, Modeling Approaches, and Practical Results

Introduction Meituan’s mission is to help people eat and live better. Its homepage search serves tens of millions of users daily and supports a growing set of retail business lines (flash‑sale, grocery, curated selection, group‑buy, etc.). The paper introduces the challenges of ranking items from multiple business domains and shares the exploration and practice carried out by the Meituan Search team.

Challenges of Multi‑Business Ranking The main difficulties are: (1) large heterogeneity in supply and fulfillment across businesses, making a unified mixed‑ranking model hard to train; (2) imbalanced data distributions because of different traffic volumes; (3) the need to jointly optimize business ordering and block size in the aggregated display, which cannot be solved by independent optimizations.

Mixed‑Tower Modeling A five‑tower architecture is built, one sub‑tower for each business (flash‑sale, grocery, take‑away, curated selection, group‑buy). An ESMM‑style component is added to better learn order information. Offline metrics improve noticeably and online A/B tests show a 20‑bp increase in click‑through rate (CTR) and a 37‑bp increase in purchase‑rate.

Feature Selection – Group Lasso Features from all businesses are concatenated, leading to many noisy dimensions. Group Lasso regularization groups features and applies L2 within groups and L1 across groups, yielding a sparse feature set. This version keeps offline NDCG roughly unchanged.

Feature‑Gate Mechanism To allow business‑specific importance, a learnable gate computes a weight vector for each feature via a linear layer followed by Softmax. The weighted features are then fed into the corresponding business sub‑tower. This approach improves offline NDCG by 16 bp compared with pure Group Lasso.

Probabilistic Graph Model A latent variable is introduced to model the probability of a request belonging to each business. A prior network predicts the business distribution from query, user, and context; a posterior network (used only during training) predicts it from user behavior. The training objective maximizes the Evidence Lower Bound (ELBO). This model raises offline NDCG by 39 bp and online CTR by 25 bp.

Aggregation Modeling (GSRM) When business results are displayed as blocks, both block order and block size affect user experience. The Grouping Search Results Model (GSRM) jointly predicts block position (as a CTR classification task) and block size (as a regression task) using a multi‑task MMoE backbone and incorporates user behavior sequences via self‑attention. This model improves offline NDCG by 10 bp and online CTR by 12 bp, purchase‑rate by 9 bp.

Heterogeneous Dual‑Sequence Modeling User behavior is split into item‑level and block‑level sequences. Both are encoded with self‑attention; item sequences are summed, block sequences are attention‑pooled with query, time, and location as targets. This design yields an additional 10 bp offline NDCG gain and online CTR/purchase improvements.

Block‑Size Prediction The block‑size task uses a regression loss (HuberLoss) while the preference prediction uses LambdaLoss. Multi‑gate Mixture‑of‑Experts (MMoE) shares parameters, and preference tower hidden states are transferred to the size tower (inspired by AITM). The combined model keeps offline NDCG stable and raises online purchase‑rate by 3 bp.

Summary and Outlook The paper summarizes the iterative improvements: multi‑tower structures, feature‑gate, probabilistic graph, and GSRM. Future work includes adaptive feature selection, AutoML‑driven sub‑tower sizing, automatic multi‑task loss weighting, and context‑aware mixed‑ranking.