Meituan DSP Strategy: Real‑time Bidding, Recall, CTR and Value Prediction

In recent years, online advertising, especially real‑time bidding (RTB), has become a dominant channel. A Demand‑Side Platform (DSP) receives bid requests from AdExchange, maps users via cookies or device information, and uses user profiles to filter traffic, estimate click‑through rate (CTR) and conversion value, aiming to maximize ROI.

Meituan’s DSP leverages massive in‑site user behavior to build precise user portraits and serves both external traffic and post‑click landing pages. The bidding process consists of two major steps: (1) real‑time bidding on the AdExchange request, and (2) after a winning bid, the user clicks to a second‑page (二跳页) where further browsing and conversion occur.

The DSP strategy is supported by two services: RecServer (recall) and PredictorServer (CTR/value prediction). RecServer provides candidate ads for both external and second‑page placements, while PredictorServer predicts CTR and click value for ranking and pricing.

Recall Strategies

1. Real‑time behavior recall : tracks recent clicks, favorites, purchases and decays recall probability over time.

2. Location‑based recall : includes (a) real‑time geo‑recall, (b) real‑time hot‑item recall within the current business district, and (c) preference‑based recall using historical district preferences (requires user ID).

3. Collaborative filtering : combines user‑based and item‑based similarity to expand candidate sets.

4. Scenario‑aware matrix factorization : uses FM models with features such as season, time of day, location, and weather to capture O2O consumption patterns.

CTR Estimation

Features are derived from three dimensions: user, ad, and media, plus matching and composite features. The model uses Logistic Regression with engineered features; dimensionality reduction is achieved by representing users through behavior and profile rather than raw IDs.

Negative samples are carefully selected to avoid bias from invisible ad slots, and sampling ratios are adjusted to keep the training set balanced. After sampling, predicted CTRs are calibrated back to real values using a correction formula.

Value Prediction

Two stages are described: (1) simple CVR estimation for landing‑page ads, using GBDT/FFM models; (2) regression on post‑click conversion amount for second‑page ads, ultimately adopting a GBDT model.

Evaluation & Monitoring

Offline metrics include AUC and Normalized Entropy (NE). Online validation uses A/B testing (showing 30% CTR lift for external ads and 13% for second‑page ads) and real‑time monitoring of hourly AUC and mean prediction values to detect anomalies.

The article concludes with future work such as finer traffic filtering, dynamic bid adjustment using PID controllers, and more precise location‑based recall.