Deep Learning Advances for Click‑Through Rate Prediction in Meituan's Location‑Based Advertising

In the post‑deep‑learning era, simply increasing model complexity no longer yields gains for advertising quality estimation. Meituan's in‑store advertising team tackled this by leveraging the flexibility of deep neural networks to address two core challenges of their O2O business: (1) location‑based service (LBS) distance constraints and (2) long‑term periodic user behavior.

The article first outlines the business background, describing how Meituan’s search ads, list ads, and feed ads differ in intent strength and how LBS constraints make spatial context a dominant factor. It then introduces four technical breakthroughs:

Context bias‑aware estimation via position‑wise combination: a Deep Position‑wise Interaction Network (DPIN) models multiple ad positions simultaneously, captures position bias and contextual bias, and aggregates local natural context to reduce display bias. Deployments across major ad slots increased RPM by 2‑3%.

Ultra‑long sequence modeling with spatiotemporal dependence: a pre‑trained long‑sequence encoder separates short‑term and long‑term interests, while a Spatio‑Temporal Activator Layer jointly attends to request time, request location, and merchant location. This design improved RPM by 2‑5%.

Dynamic ad‑candidate generation: candidate volume, type, and compute are dynamically adjusted based on city, business category, and traffic value, yielding up to 20% RPM lift in high‑density cities and 10‑15% improvements in heterogeneous categories.

Catastrophic forgetting mitigation and continual learning: a memory‑augmented architecture with data replay and multi‑tower training prevents the model from overwriting historic periodic patterns, delivering 2‑3% RPM gains.

All four modules are integrated into a unified network composed of Representation, Memory, and Combination units, enabling high‑precision, high‑throughput inference (from >30 ms to <1 ms for ultra‑long sequences). The paper reports that these techniques have been published in top conferences such as SIGIR and CIKM, and that they are being extended to conversion‑rate and transaction‑value prediction.

Finally, the article discusses emerging trends: dynamic inference scaling, differentiated evaluation metrics (e.g., PAUC for position bias), multi‑objective link‑level optimization, blurred boundaries between search and recommendation, and a shift toward unsupervised/self‑supervised learning in advertising systems.