Graph Computing for Financial Credit Risk Control and Anti‑Fraud: Architecture, Challenges, and Lessons Learned

Background Introduction

The discussion starts with the business background of credit lending, highlighting the rapid development of AI and big‑data technologies that have driven the financial credit industry toward intelligent, digital operations.

Credit‑Related Terminology

Credit: the amount of credit granted to a user, enabling borrowing on the platform.

Order: installment purchase, where the user pays the principal and interest over multiple periods.

Credit‑Lifecycle: pre‑loan, in‑loan, and post‑loan stages; the focus here is on early fraud detection.

New Customer: a user without a complete repayment history, typically higher risk.

Old Customer: a user with at least one full repayment cycle, generally lower risk.

Data Discrepancy: differences between credit and order data, affecting timeliness and richness.

Graph Model Stakeholders

In Akulaku, graph algorithm engineers mainly collaborate with anti‑fraud business personnel, forming two stakeholder groups:

Technical staff: model analysts and engineers who explore new technologies.

Business staff: risk‑control strategists who ensure stable model performance.

Applications of Graph Computing in Financial Risk Control

Graph computing supports two major anti‑fraud use cases:

Gang (cluster) detection: identifying members of fraudulent groups, extracting group features, and building models for automatic detection.

Association discovery: analyzing topological structures and abnormal patterns to construct encoding and modeling pipelines.

Practical constraints include limited data availability in certain business stages, differing timeliness requirements (order stage demands higher speed than credit stage), and the need for fast graph computation within strict latency budgets.

Requirements for Graph Computing Systems

Stability : both technical stability of services and business stability of model scores.

Timeliness : meet latency targets (e.g., 500 ms response in the order stage).

Accuracy : ensure online features match offline back‑testing results, avoiding data leakage.

Controllability : provide explainability and verifiability, with clear feature validation.

Evolution of Graph Computing Architecture

Stage 1 – Initial Graph Mining

Implementation used separate offline and real‑time pipelines. Offline algorithms performed gang and feature mining with a T+1 update cycle, while real‑time rules (e.g., black‑list checks) relied on a graph database for immediate queries.

Key challenges: limited data coverage, high latency for real‑time needs, and difficulty handling sliding time windows for incremental updates.

Stage 2 – Real‑Time Graph Mining

The second stage introduced incremental graph clustering (based on Louvain) to move gang detection into the credit stage, achieving 100 % coverage and high availability of the graph database.

Feature computation shifted to an event‑driven approach using PolarDB for intermediate tables, enabling precise back‑tracking and validation of real‑time features.

Stage 3 – End‑to‑End Graph Modeling

The current stage incorporates graph convolutional networks for end‑to‑end fraud detection, leveraging real‑time data from a streaming warehouse and online inference on the graph database.

Experience Summary

Stability

Focus on database selection, high‑availability master‑slave setups, and comprehensive monitoring.

Timeliness

Implement real‑time graph mining algorithms and asynchronous feature computation to meet sub‑second latency requirements.

Accuracy

Maintain strict feature back‑testing pipelines to ensure online and offline consistency, using them also to detect data‑quality anomalies.

Controllability

Adopt a progressive modeling strategy: start with simple rule‑based models, evolve to interpretable gang‑feature models, and finally to deep end‑to‑end models, ensuring thorough validation at each step.

References:

Li, X. and W. Zhang. “HGsuspector: Scalable Collective Fraud Detection in Heterogeneous Graphs.” (2018).

Alexandre Hollocou et al. “A Streaming Algorithm for Graph Clustering.” NIPS 2017 Workshop.