Algorithmic Practices for Meituan's Content Intelligent Distribution

This article, derived from Meituan Tech Salon Session 80, presents the algorithmic practice of content intelligent distribution on the Meituan platform. It is organized into three parts: (1) the current status and challenges of content search in the Dazhong Dianping (大众点评) ecosystem; (2) practical optimizations applied across the entire search pipeline; and (3) a summary with future outlook.

1. Current Status and Challenges – Dazhong Dianping has accumulated massive multimodal (text, image, video) user‑generated content. Users heavily rely on search to discover local services, making content search a key decision‑support tool. Compared with typical web, e‑commerce, or merchant search, content search faces unique difficulties: mixed structured and unstructured supply, huge and frequently updated content volume, sparse user‑behavior signals, the need to balance consumption metrics with relevance, and heterogeneous result mixing with merchants and deals.

2. Content Search Optimization Practices

2.1 Supply Understanding – Explicit labeling (category, fine‑grained thematic, attribute tags) and implicit multimodal representation via a self‑trained multimodal pre‑training model that aligns image‑text features in a unified space.

2.2 Recall Stage – Three recall strategies are employed: semantic recall (multi‑granular semantic units), personalized recall (geographic preference, historical similarity), and strategy‑based recall (latest, hot, recipe‑type, author‑specific).

2.3 Ranking Stage – Includes coarse‑ranking, fine‑ranking, multi‑objective fusion, and heterogeneous mixing. Techniques such as representation distillation, score distillation, sequential distillation, cross‑tower features, and multi‑gate networks (MMoE + PPNet) are used. Input representations cover query semantics, user sequence modeling with zero‑vector attention, multimodal embeddings, and dynamic feature importance modeling. Multi‑objective modeling addresses the trade‑off among click, dwell time, and interaction metrics using MMoE, PPNet, and specialized loss weighting. Output layer introduces exploration structures and calibrated learning objectives (LambdaLoss + LogLoss) to improve score stability.

2.4 Satisfaction Optimization – Beyond consumption metrics, relevance, freshness, locality, and content quality are modeled. Automated labeling pipelines are explored to reduce manual annotation cost while maintaining acceptable accuracy.

2.5 Multi‑Objective Fusion – Precise score estimation, AutoML‑driven hyper‑parameter search, and distribution control factors (new content boost, old content regulation, proximity‑based promotion) are integrated.

2.6 Heterogeneous Mixing – Content results are mixed with merchants, group‑buy deals, etc., using end‑to‑end, value‑fusion, and sequence‑generation approaches. Traffic patterns (fast vs. slow decision scenarios) are considered to dynamically allocate compute resources.

3. Summary and Outlook – The optimizations have significantly increased content search penetration and user satisfaction. Future work includes automated issue detection, broader adoption of large language models across the pipeline, and further improvements in timeliness and performance to strengthen Meituan’s local lifestyle community.

The article also lists the authors (Yi‑Fan, Tao‑Ran, Tao‑Feng, Sheng‑Yu) and provides a set of academic references covering ranking, personalization, and calibration techniques.