DATA LAYER
Multi-Source Data
Acquisition and Fusion technology
E-commerce data: shopping history, spending patterns, and delivery address stability.
Social data: Social networks, behavioral preferences (compliance considerations
required).
Operator data: call duration, network usage duration, and recharge habits.
Behavioral data: App usage duration, clickstream, and page dwell time
Device data: Device fingerprint, IP address, GPS location (for anti-fraud purposes).
Third-party data: Blacklist information, cross-borrowing records, and judicial enforcement records from professional data service providers.
Big Data Processing & Storage
Distributed storage solutions such as HDFS and HBase are designed to store massive amounts of structured and unstructured data.
Distributed computing technologies such as Spark and Flink enable high-speed ETL(Extract, Transform, Load) processing and real-time computation for massive
datasets.
Knowledge Graph Technology
This serves as the “brain” of the data layer. It constructs a vast relational network by integrating entities from various sources (such as individuals, enterprises, devices, phone numbers, addresses, etc.) and their interrelationships.
Core application: Identification of complex gang fraud. For example, through knowledge graphs, multiple seemingly unrelated applicants are discovered who actually share the same device, IP address, or contacts, thereby exposing fraudulent gangs.
ALGORITHM Layer
Deep Learning Models
Neural networks: Capable of automatically learning complex nonlinear features, particularly excelling in processing unstructured data such as images, text, and
sequence data.
Application scenarios ：
Image recognition: Automatic identification and verification of the authenticity of ID cards, bank cards, and business licenses.
Natural Language Processing: Analyze user-customer service conversation texts to detect fraudulent intent or credit risks.
Sequence model: Analyzes user transaction behavior patterns to detect abnormal transaction patterns in real time.
PLATFORM LAYER
Real-Time Computing Engine
Risk decision-making often occurs at millisecond-level precision. By leveraging technologies such as Flink and Storm, real-time processing of user behavior streams and transaction flow data can be achieved, enabling instant model invocation for real-time scoring and decision-making. For instance, fraud risk assessment can be completed in the instant a user makes a card payment.
Rule Engine
The rule engine can be integrated with machine learning model scoring to form a hybrid decision-making system of ‘rules + models’.
Feature Platform
The process of transforming data into model-ready ‘features’ is standardized and platformized.
It manages thousands of features, ensuring consistency during model training and online prediction, thereby significantly enhancing the efficiency and reliability of model iterations.
Model Lifecycle Management (MLOps)
Responsible for model deployment, monitoring, iteration, and decommissioning. The financial world is dynamically evolving, and models may become “ineffective.” MLOPS technology is employed to monitor the degradation of model performance
and automate the retraining and deployment of models.
APPLICATION LAYER
Value Output of Risk Control
Credit Risk
Core technologies: machine learning scoring card, big data credit assessment.
Objective: To assess the borrower’s default probability for credit approval, credit limit pricing, and post-loan management.
Fraud Risk
Core technologies: Device fingerprinting, biometric recognition, behavioral sequence analysis, knowledge graph, and real-time decision engine.
Objective: To detect malicious activities including application fraud, transaction fraud, cash-out, and fake transactions.
Market and Operational Risk
Core technologies: big data monitoring, NLP (for public sentiment and news analysis), and stress testing models.
Objective: To monitor liquidity risks and identify internal non-compliant operations.