Deep Learning Model Applications and Optimizations for Recommendation Ranking at Meituan

This article introduces the use of deep learning models for solving information overload and uncovering user demand in Meituan's extensive lifestyle e‑commerce platform, particularly in the "Guess You Like" recommendation scenario.

Meituan collects billions of daily user‑item interaction logs, providing massive training samples. To handle this scale, a distributed training framework (PS‑Lite) is employed to train Multi‑task DNN models.

Feature engineering covers user, item, context, and behavior features. Continuous features are processed with equal‑frequency (quantile) binning to achieve near‑uniform distributions, and low‑frequency sparse features are filtered out.

The Multi‑task DNN shares early layers between click‑through‑rate (CTR) and order‑conversion (CVR) tasks, with separate heads for each. Additional layers include a Missing Value Layer that learns adaptive imputations for absent continuous features, and a KL‑divergence bound that enforces consistency between predicted CTR × CVR and the observed order probability.

Further model enhancements incorporate Neural Factorization Machines (NFM) for low‑order feature interactions and various user‑interest vector constructions (average, max, weighted pooling with attention). Experiments show faster convergence with NFM but limited AUC gain.

Training efficiency is improved by serializing pre‑computed metadata to reduce parsing overhead and by introducing a worker‑interrupt mechanism that skips lagging workers after most have finished an epoch, thus shortening overall training time.

Offline evaluations demonstrate significant AUC improvements over baseline XGBoost, and online A/B tests report notable CTR lifts after deploying the Multi‑task DNN in the "Guess You Like" feed.

The authors conclude with future directions: deeper business rule integration via multi‑task objectives, richer context feature exploration, and continued network architecture innovation.