Financial Fraud Detection Overview

Transaction fraud is a forty-three billion dollar problem annually. It poses a major challenge for financial institutions that struggle to detect and prevent increasingly complicated fraudulent activities. Traditional fraud detection methods that rely on rules-based systems or statistical methods are reactive and increasingly ineffective in identifying elaborate fraudulent activities. As data volumes grow and fraud tactics evolve, financial institutions need more proactive, intelligent approaches to detect and prevent fraudulent transactions.

The Financial Fraud Training container provides capabilities to train Graph Neural Network (GNN) based XGBoost models to predict fraud scores of credit card transactions. Based on user-provided training configuration, the container first builds a GNN model that produces embeddings for credit card transactions, and then it uses the transaction embeddings to train an XGBoost model to predict the fraud scores of the transactions. The container encapsulates the complexity of creating the graph in cuGraph. Once the graph is created, the GNN model is trained and used to produce the embeddings that are then fed to XGBoost. It also provides the option to train standalone XGBoost models on input data.

Architecture

Fraud detection workflow in a typical payment processing environment can be broken down into three steps:

1. Data Preparation

2. Model Building

3. Inference

The Financial Fraud Training container provides the capabilities for model building (step 2) in this workflow.

Try It Out

Users can try out Financial Fraud Detection at NVIDIA.