Deep Learning and Ranking Model Evolution for Hotel Search at Meituan

This article, based on a 2018 QCon talk by Meituan senior engineer Zhai Yitao, describes the application of deep learning to hotel search NLP and the evolution of ranking models in Meituan's hotel platform.

Business characteristics : Hotel search is an O2O service where users search for hotels online and stay offline. It differs from web or e‑commerce search in that location constraints and transaction intent are critical. The system consists of three modules: retrieval (query understanding and recall), ranking (machine‑learning based ordering), and business rules.

Retrieval and intent understanding : Various query issues such as landmark terms, structured queries, cross‑city queries, and synonyms are handled by a multi‑stage pipeline. Landmark queries are detected via NER, mapped to coordinates, and retrieved by geo‑search. A multi‑level retrieval architecture (basic, secondary, core‑term, and cross‑city retrieval) reduces no‑result rates.

Ranking model evolution :

Initial model: XGBoost (GBDT) leveraging abundant continuous features (price, distance, rating).

MLP stage: Fully‑connected neural networks with 3‑6 layers and pyramid architecture (e.g., 1024‑512‑256).

Embedding of discrete features: FNN, DeepFM, PNN, DCN were explored; only FNN reached production.

Wide&Deep with multi‑task learning (CTR and CVR) became the current online model, enriched with billions of engineered cross‑features.

Challenges : Data sparsity (hotel stays are low‑frequency), diverse business scenarios (domestic, overseas, long‑term, hourly rentals), complex user contexts (local vs. cross‑city, business vs. tourism), and supply constraints (limited room inventory).

Model serving : The team migrated from ad‑hoc TensorFlow scripts to TF‑Serving and finally to the in‑house MLX platform, enabling seamless offline‑online pipelines and real‑time feature updates.

Technical roadmap : Different business stages (startup, early growth, stable, bottleneck) dictate appropriate model complexity—from simple heuristics to sophisticated deep ranking models.

Conclusion : Hotel search requires both relevance and personalized ranking. By modularizing retrieval, intent understanding, and ranking, and by iteratively adopting more advanced AI techniques, Meituan achieved significant improvements in user purchase experience while keeping the model stack aligned with business needs.