Edge AI Re‑ranking in Meituan/Dianping Search: Architecture, Algorithms, and Deployment

Edge intelligence refers to running AI applications on mobile devices. This article presents the practice of deploying large‑scale deep learning models for search re‑ranking on the client side of the Meituan/Dianping app.

1. Introduction – With the rapid growth of big data and AI, cloud computing can no longer meet the latency and privacy requirements of some scenarios. Deploying AI on the edge reduces data transmission, protects user privacy, and enables real‑time personalization. Major tech companies (Google, Apple, Alibaba, ByteDance, etc.) have already released edge AI solutions; Meituan/Dianping applies the same idea to its search service.

2. Why On‑Device Re‑ranking? – Cloud‑side ranking suffers from two main drawbacks: (1) delayed list updates due to pagination, and (2) minute‑level latency of real‑time feedback signals processed by streaming platforms (Storm, Flink). On‑device re‑ranking can refresh results within a page, sense user feedback instantly, and keep data local for privacy.

3. On‑Device Re‑ranking Algorithms

3.1 Feature Engineering – The feature set mirrors the cloud pipeline (User, Item, Query, Context) with additional edge‑specific real‑time feedback features. Table 1 in the original text lists basic, bias, and feedback features.

3.2 User Feedback Sequence Modeling – Various sequence models (DIN, DIEN, BST) are explored. A Deep Feedback Network (DFN) splits the feedback into positive (click) and negative (exposure‑no‑click) sequences and applies cross‑attention. To improve the signal‑to‑noise ratio of negative feedback, long‑exposure events are weighted more heavily.

3.3 Multi‑View Feedback Attention Network (MVFAN) – Extends DFN by adding multiple view embeddings (category, price, distance, etc.) and applying multi‑head attention across these views.

3.4 Re‑ranking Model Design – Context‑aware list‑wise models and Transformer‑based encoders are used to capture interactions among the top‑N candidates. The final model outputs a score for each candidate, optimized with ListWise LambdaLoss and ΔNDCG to emphasize top‑position clicks.

4. System Architecture & Deployment Optimization

4.1 Architecture – The edge re‑ranking system consists of three modules: (1) intelligent trigger module, (2) on‑device re‑ranking service (feature processing + inference engine), and (3) native post‑processing for result insertion.

4.2 Model Deployment – Because mobile storage is limited, the model is split into a dense network (≈10 MB) and a large ID‑embedding table (≈80 % of parameters) served from the cloud. Sparse embedding + dense network are combined at inference time.

4.3 Model Compression – A Meituan‑specific compression tool reduces the dense network to <1 MB with negligible accuracy loss and no noticeable power increase.

4.4 Training & Estimation Platform – An end‑to‑end platform (Augur/Poker) integrates feature processing, model training, and online A/B testing for edge models, streamlining the iteration cycle.

5. Experimental Results

Offline experiments (Table 2) show consistent gains in CTR and QV_CTR. Online A/B tests report a 0.25 % increase in main‑search QV_CTR and a 0.43 % increase in the food‑channel QV_CTR, with improved click rates on later pages.

6. Conclusion & Future Work

The edge re‑ranking solution successfully improves search relevance, reduces latency, and protects privacy. Future directions include federated learning for privacy‑preserving model updates, richer trigger strategies, more robust negative‑feedback modeling, and personalized “thousand‑model” deployments.