Design and Implementation of Ctrip's User Precise Marketing System

1. Introduction

Ctrip's financial core products include installment payment, cash loan, co‑branded credit cards, and wealth management, along with smaller services such as flash travel cards and virtual coins. Recommending these innovative products to interested users is challenging, prompting the development of a user precise marketing system by the payment data team.

2. System Design Goals

1) Non‑intrusive and decoupled from business systems; business sides only need to provide standard RESTful interfaces.

2) Flexible marketing page deployment, supporting both common and personalized strategies that can be adjusted in real time without developer involvement.

3) Support multiple delivery modes such as priority‑based, random, and model‑driven, with model effectiveness quantified via AB testing.

4) Reporting and monitoring, including real‑time traffic usage, clicks, offline conversion, API performance, and offline data collection for strategy effect tracking.

2.1 Overall Architecture

1) Strategy Management Module: collects business requirements and consolidates rules, focusing on traffic shunting and model‑based scoring to improve conversion.

2) Model Training Module: defines black‑white samples per business line and uses machine‑learning algorithms to score user eligibility.

3) Data Collection/Analysis Module: builds comprehensive offline and real‑time monitoring, logs events to Hive, and analyzes conversion at each step to optimize strategies.

2.2 Interaction Sequence

Business interfaces are intercepted to display marketing pages; a thread pool processes rules per popup type, and response timeouts are enforced to maintain system stability.

3. Core Functions

The system focuses on traffic usage rate and conversion rate. Traffic usage is refined through shunting, while conversion is enhanced via machine‑learning models.

3.1 Traffic Usage Rate – Shunting

Shunting allocates limited traffic based on strategies such as random uniform, priority‑based, or recommendation‑algorithm‑based distribution.

3.1.1 Random Uniform

Ensures even distribution across business lines and consistency for the same user using hash modulo or MurmurHash3 for high concurrency.

3.1.2 Predicting Daily Traffic

When multiple business lines intercept a user, one is randomly chosen; to avoid unfair allocation, daily traffic is forecasted using time‑series models to set budgets.

3.2 Conversion Rate – Model

3.2.1 Model Y‑value Definition

The target is user account opening rather than click, because redesigns affect click rates but not overall conversion.

3.2.2 Model Development and Deployment

After testing several algorithms, GBDT + LR was selected; with hundreds of features, daily full‑user scoring is performed on a Spark‑ML Hadoop platform.

3.2.3 Score Standardization

Scores from different lines are normalized to a common baseline using segmented conversion rate calculations and min‑max scaling.

4. Strategy Considerations and Solutions

4.1 Model Threshold with Low Traffic Budget

When a line receives minimal traffic, raising the model threshold can prioritize high‑quality exposure and prevent early depletion of low‑quality traffic.

4.2 Coexistence of Modeled and Non‑Modeled Lines

For lines without models, random uniform selection is used to merge them with modeled lines during page delivery.

4.3 New Business Line Onboarding

New lines start with allocated traffic to gather samples before model scoring.

4.4 Quantifying Model Effect – AB Test

AB tests evaluate conversion impact; in credit‑type scenarios, AB testing may be less suitable, so target metrics like opening conversion are visualized directly.

4.5 Achieving Near 1:1 AB Test Traffic

Embedding tracking points in A/B samples and adjusting model scores offline helps balance traffic distribution.

5. Conclusion

The marketing system leverages Ctrip user data and machine‑learning algorithms to perform flexible segmentation and recommendation, addressing traffic utilization and conversion challenges, while supporting real‑time strategy adjustments and AB‑test‑driven optimization.

Data engineers can trace end‑to‑end conversion through comprehensive backend and UBT logs, enabling continuous strategy refinement.