Deep Learning Applications in Meituan‑Dianping Recommendation System

Meituan‑Dianping, the largest domestic lifestyle service platform, faces rapid growth in users and merchants across diverse business scenarios (food, accommodation, travel, etc.). To improve user experience, the company continuously optimizes its recommendation algorithms, aiming for high precision, richness, and user delight.

The paper introduces the recommendation system architecture, highlighting challenges such as business form diversity and varying user consumption contexts. A two‑stage pipeline—recall and ranking—is employed, where recall generates candidate items using multiple strategies (user‑based, model‑based, item‑based, query‑based, location‑based) and ranking fuses these candidates.

Deep learning is applied to address limitations of linear models (e.g., over‑reliance on historical clicks). The authors adopt Google’s 2016 Wide & Deep framework, combining a linear “wide” component for memorization with a deep neural network for generalization. Feature engineering includes user, item, and context profiles, extensive cross‑features, and normalization (Min‑Max). Advanced feature transformations such as fast aggregation and hierarchical feature combinations are also discussed.

Various optimizers are evaluated: SGD, Momentum, Adagrad, and Adam. Adam is selected for its balance of adaptive learning rates and momentum, providing stable convergence. Loss functions are compared, with Cross‑Entropy chosen over Mean Squared Error due to better gradient behavior for CTR prediction.

The final Wide & Deep model consists of a wide linear part handling sparse cross‑features and a deep part with three ReLU layers followed by a sigmoid output. Training uses TensorFlow/Keras, batch size 50 000, 20 epochs, and over 70 million training samples. Offline and online A/B tests demonstrate consistent CTR improvements and better novelty in recommendations compared to baseline models.

Future work includes integrating RNNs to capture temporal dynamics and exploring reinforcement learning for context‑aware recommendation policies.