Algorithmic Practices for Meituan's Content Intelligent Distribution

Meituan's local life platform has accumulated massive user‑generated text, image, and video content, enabling richer user needs beyond transactions such as pre‑purchase discovery and post‑experience sharing, turning Dazhong Dianping into a community for local food, entertainment, and travel.

Content search is a key tool for decision‑making, presenting diverse results (e.g., merchants, deals, notes, videos) when users query terms like "hotpot" or "Christmas events," thereby enhancing community engagement.

Compared with web, e‑commerce, and merchant search, content search differs in optimization goals (balancing satisfaction, clicks, dwell time, interactions), strong LBS constraints, varied supply types (notes, reviews, guides, recipes), semi‑structured data, larger supply scale, and faster update cycles.

The system faces four main challenges: (1) coexistence of structured and unstructured multi‑type supply, (2) massive, frequently updated content causing sparse user behavior and strong regional constraints, (3) need to jointly improve consumption metrics and search satisfaction, and (4) heterogeneous mixing of content with merchants and deals in final results.

Supply Understanding : Explicit tagging includes category, fine‑grained (topic, concept), and attribute tags (e.g., policy‑related, duplication). Implicit representation is achieved via a self‑developed multimodal pre‑training model that aligns image‑text features using contrastive loss, masked learning, and image‑text matching.

Recall Stage : Semantic recall (multi‑granular semantic units), personalized recall (geographic and historical preferences), and strategy recall (latest, hot, author‑specific, recipe‑specific) are employed, with semantic and personalized recall largely implicit.

Ranking Stage : Includes coarse‑ranking, fine‑ranking, multi‑objective fusion, and heterogeneous mixing. Coarse‑ranking balances accuracy and latency using user‑wide behavior samples and representation/distillation techniques. Fine‑ranking enhances query, user, document, and context representations via query semantics, user sequence modeling with zero‑vector attention, multimodal vectors, and dynamic feature importance modeling (e.g., EPNet, MaskNet). Multi‑objective modeling uses MMoE and PPNet with dropout and skip connections to mitigate gate polarization, and gradient control for sparse embeddings. Output layer introduces exploration structures (cold‑start, adversarial gradient) and calibration loss (LambdaLoss + LogLoss) for stable score estimation.

Satisfaction Optimization : Beyond consumption metrics, relevance, freshness, locality, and content quality are evaluated. Automated labeling pipelines replace costly manual annotation, achieving comparable accuracy while reducing expense and improving stability.

Multi‑Objective Fusion : Precise score estimation, automated hyper‑parameter search for Pareto‑optimal trade‑offs, and distribution control (new content boost, old content governance, proximity‑based promotion) are applied.

Heterogeneous Mixing : Content results are mixed with merchants and deals using end‑to‑end models, value‑fusion formulas, and sequence generation, considering time‑of‑day consumption patterns (e.g., lunch vs. night‑snack) and dynamic resource allocation.

In conclusion, Dazhong Dianping's content search has significantly improved user engagement and growth by building a robust content‑centric search pipeline, addressing supply challenges, and continuously optimizing each link of the system, with future work focusing on automated issue detection and large‑model integration.