Application of Graph Technology in Financial Anti‑Fraud

01 Graph Technology Background in Financial Anti‑Fraud

The financial credit sector faces increasingly sophisticated fraud that has evolved from isolated incidents to large‑scale, organized schemes, creating a massive black‑industry chain that disrupts normal financial order and pressures risk‑control teams.

Challenges include rapid data growth, complex data formats, and the need for early‑stage detection to prevent large‑scale losses while balancing risk mitigation with customer experience.

02 Role of Graphs in Financial Anti‑Fraud

Graphs address two key fraud characteristics: extreme class imbalance (few fraudulent samples) and strong clustering of malicious actors. Graph‑based community detection and label propagation overcome data sparsity without relying on balanced samples.

Graphs also capture relationships among entities (e.g., phones, merchants) that single‑user features miss, enabling detection of organized fraud rings and hidden risk patterns.

03 Graph‑Based Anti‑Fraud Case Study

By integrating multi‑dimensional data (features, behavior, funds), heterogeneous sources are unified into various graph assets such as media‑network graphs, merchant‑relation graphs, and fund‑relation graphs. Community‑mining algorithms identify tightly‑connected clusters, which are scored against known black entities to flag high‑risk groups for intervention.

04 Graph‑Driven Perception, Judgment, Decision, and Disposal

The risk‑control workflow is divided into four layers:

Perception: Real‑time sub‑graph processing (hundreds of billions of edges daily) detects potential risks such as abnormal fund flows.

Judgment: Experts analyze known black‑user cases on a visual graph platform, summarizing common fraud patterns and manually labeling data.

Decision: Perception results and expert‑derived patterns become graph features applied in decision‑making policies.

Disposal: Based on decisions, targeted actions (black‑listing, credit limit reduction, transaction‑level interception) are executed.

05 Evolution of Graphs in Financial Anti‑Fraud

Early stage relied on traditional statistical features and rule‑based models. Introducing graphs enabled expert‑driven pattern matching (e.g., fund loops, abnormal flows) to precisely target risky cohorts.

Subsequent stages moved from individual targeting to gang‑level control using graph‑learning algorithms that aggregate temporal and spatial information.

Later, a full‑stack risk‑graph platform built on Ant Group’s TuGraph (PhStore, Geabase, Geaflow) provides storage, real‑time query, and offline computation for trillion‑scale graphs, offering a unified pipeline from data modeling to deployment.

To reduce expert dependence, an automated graph risk‑pattern mining project was launched, employing graph extraction, noise filtering, mining, matching, and quantitative evaluation. Advanced algorithms (pre‑pruning, heuristic search, GraphPi for isomorphism) enable large‑scale, high‑performance mining, recognized with awards and academic publications.

06 Summary and Outlook

Graphs now support the entire anti‑fraud lifecycle: pre‑risk perception, mid‑process interception, and post‑risk monitoring. Their abilities to enrich samples, elevate data dimensions, and provide strong interpretability make them a decisive tool in fraud prevention.

Future work aims to combine AI with graph techniques to discover unknown, low‑visibility risks, achieving proactive defense and a cleaner financial ecosystem.

Thank you for your attention.