Deep Learning Practices in Meituan O2O Service Search

Meituan provides comprehensive life services and its search system supports about 40% of platform transactions, handling millions of POIs and billions of SPUs with daily query volumes at the hundred‑million level.

The search service faces challenges like heterogeneous business data, non‑standardized merchant attributes, and dynamic user demands that vary by location, time, and weather.

To address these, Meituan focuses on intelligent matching and multi‑dimensional retrieval, distinguishing explicit and implicit user intents and performing component analysis to extract address and category information from queries.

Business recognition relies on a knowledge base and data‑driven iterative feedback, employing CRF, BiLSTM‑CIF, and reinforcement‑learning techniques to improve entity extraction and intent detection.

Semantic modeling includes a customized DSSM twin‑tower architecture, RNN‑based session modeling, and vectorized representations of queries, POIs, and users for efficient recall.

Personalized ranking combines a coarse recall layer, click‑through and conversion prediction models, and business‑rule‑based ordering, evolving from linear models to GBDT, Pairwise, real‑time, and deep learning models.

Real‑time feature extraction caches user behavior streams in Storm, deriving preferences for category, price, and distance, and feeds them into online models.

The deep learning framework supports massive offline training, streaming updates, and online prediction, utilizing MLP, embedding learning, Wide&Deep, PNN, DeepFM, and DCN models to handle sparse features and high‑order interactions.

Future work aims to deepen intelligent matching through richer component analysis and multi‑objective optimization, and to enhance ranking models with user‑interest modeling similar to DIEN, integrating textual, categorical, and contextual signals.