Evolution of Ctrip Financial Risk Control Models: From Data Platform to AI‑Driven Scoring and Anti‑Fraud Systems

The presentation by Zeng Fanxiang, leader of Ctrip Financial's big data team, outlines the complete evolution of the company's risk control algorithms, using a time‑based narrative that moves from sparse data and coarse features to abundant data, refined features, increasingly complex models, and markedly improved performance, with a focus on application scoring and anti‑fraud models.

Business Introduction : Ctrip Financial operates three core consumer‑finance modules—consumer installment ("拿去花"), cash installment ("借去花"), credit cards, and supply‑chain finance.

Data Platform : The data‑mid‑platform is abstracted into three layers: a foundational data layer, a business‑abstraction model layer, and an algorithm‑model layer. Its goals are to reduce communication costs, improve collaboration, provide high‑availability data, and support user profiling, feature engineering, and model services.

Risk Model System : Risks are divided into controllable (fraud, credit, operational) and uncontrollable (market, systemic) categories. The risk‑control models span the entire customer lifecycle—pre‑loan (application scoring, A‑card), in‑loan (behavior scoring, B‑card), and post‑loan (collection scoring, C‑card). Each model serves specific business functions such as approval, pricing, and collection.

Pre‑Loan Credit Risk Model (A‑card) : This model addresses three tasks—anti‑fraud identification, credit rating, and risk pricing. Its evolution progressed from rule‑based systems to logistic regression, then to GBDT/RF/XGBoost, followed by DNN, fractal‑network architectures, and finally migration‑learning frameworks, with each stage improving predictive power.

Model Evaluation and Monitoring : Evaluation is performed on three levels: (1) machine‑learning metrics (KS, AUC, OOT testing), (2) risk‑control metrics (bucket‑wise ranking performance), and (3) business metrics (approval rate, delinquency, revenue impact). Corresponding monitoring tracks model stability, feature distribution, and drift.

In‑Loan Anti‑Fraud Model : Fraud detection operates at user‑level (coarse) and transaction‑level (fine) granularity. Challenges include long‑tail fraud distribution, adversarial behavior, and mimicry of normal activity. The solution combines traditional machine‑learning models with a social‑network‑based detection framework.

Social Network in Risk Control : A heterogeneous graph containing billions of vertices and edges is constructed from Ctrip's ecosystem. Community detection algorithms (LPA, improved Louvain) identify suspicious clusters, providing fraud indicators that now cover about 80% of loan requests and enhance collection efficiency.

Author and Recruitment : The speaker, Zeng Fanxiang, holds a Ph.D. from Beijing University of Posts and Telecommunications and was a visiting scholar at McGill University. The article concludes with recruitment notices for senior AI algorithm engineers and senior big‑data engineers, outlining responsibilities, qualifications, and contact information.