Optimizing Pre‑Ranking in Meituan Search: Knowledge Distillation and Neural Architecture Search

Pre‑ranking is a crucial module in industrial large‑scale search, recommendation, and advertising systems. Meituan’s search ranking team has explored and practiced ways to improve pre‑ranking effectiveness while respecting strict performance constraints. This article introduces the evolution roadmap of Meituan’s pre‑ranking, and presents optimization work based on knowledge distillation and automatic neural network architecture search, aiming to inspire practitioners working on similar problems.

1. Introduction In cascade ranking architectures widely used in industry, the ranking pipeline consists of recall → pre‑ranking → fine‑ranking → re‑ranking. Pre‑ranking sits between recall and fine‑ranking, filtering thousands of candidate items down to a few hundred for the fine‑ranking stage. The main challenges of Meituan’s pre‑ranking layer are sample selection bias, the need for fine‑ranking‑pre‑ranking interaction, and strict latency constraints.

2. Pre‑ranking Evolution The technology has progressed through several stages:

2016: Linear weighted combination of relevance, quality, and conversion features.

2017: Simple logistic regression (LR) pointwise model.

2018: Dual‑tower model with query/user/context and merchant embeddings.

2019: Fusion of dual‑tower outputs with GBDT to capture cross features.

2020‑present: End‑to‑end neural network (NN) models, continuously iterated.

Current industrial pre‑ranking models include dual‑tower approaches (e.g., Tencent, iQIYI) and interactive NN models (e.g., Alibaba). The following sections detail Meituan’s optimization efforts during the transition to NN‑based pre‑ranking.

3. Pre‑ranking Optimization Practices

3.1 Fine‑ranking‑linked Effect Optimization Because pre‑ranking models are simpler and use fewer features than fine‑ranking models, their effectiveness lags. To bridge this gap, Meituan applied knowledge distillation, treating the fine‑ranking model as a teacher and the pre‑ranking model as a student.

3.1.1 Teacher‑Student Result List Distillation Positive and negative samples are augmented with items from the fine‑ranking result list. Three strategies were evaluated:

Strategy 1: Add a small number of low‑rank, non‑exposed fine‑ranking items as negative samples (Recall@150 + 5 PP, online CTR + 0.1%).

Strategy 2: Randomly sample from the fine‑ranking list and use the rank position as a label to construct pairwise training data (Recall@150 + 2 PP, online CTR + 0.06%).

Strategy 3: Bin fine‑ranking positions into intervals, create interval‑based labels, and form pairwise samples (Recall@150 + 3 PP, online CTR + 0.1%).

3.1.2 Prediction‑Score Distillation A two‑stage distillation framework aligns the pre‑ranking score distribution with that of the fine‑ranking model using a mean‑square error loss weighted by a hyper‑parameter λ. This yields Recall@150 + 5 PP and online CTR + 0.05%.

3.1.3 Feature‑Representation Distillation Traditional KL‑based distillation fails to capture structured, highly correlated features. Meituan introduced contrastive learning: positive pairs are representations from the same rank bin, while negative pairs come from different bins. An InfoNCE loss maximizes mutual information between teacher and student representations, achieving Recall@150 + 14 PP and online CTR + 0.15%.

3.2 Joint Effect‑Performance Optimization Pre‑ranking must meet latency requirements while improving effectiveness. Simple DNN + distillation suffers from limited feature richness and fixed architecture. Meituan therefore employed Neural Architecture Search (NAS) combined with efficiency modeling to automatically select the best feature set and network structure under a latency budget.

The NAS framework builds on ProxylessNAS, introducing differentiable feature masks (Bernoulli‑distributed) and MixOp layers representing candidate network units (e.g., hidden sizes 1024, 512, 256, 128, 64, and 0). Efficiency modeling captures both feature retrieval latency and model‑inference latency, allowing a differentiable loss that balances accuracy and latency. The final multi‑objective loss combines the fine‑ranking teacher loss, the student loss, and weighted efficiency terms.

Experimental results show offline Recall@150 + 11 PP and online CTR + 0.12% without increasing latency. The work has been accepted by KDD 2022.

4. Summary Since 2020, Meituan has deployed MLP‑based pre‑ranking models and continuously iterated them. By leveraging three distillation techniques (result list, prediction score, and representation), applying contrastive learning for structured knowledge transfer, and integrating NAS‑driven joint effect‑performance optimization, the pre‑ranking layer has achieved significant gains without extra latency. Future directions include multi‑objective modeling for pre‑ranking and system‑wide dynamic compute allocation.