Resources & FAQ

Q: Who should use it?

A: Customers who need to build risk control system based on business data using AI technology.

Q: What customer experience is like?

A: You can deploy the solution into your own Amazon Web Services account via the CloudFormation template in a few minutes. All necessary resources will be partitioned well. Please follow the deployment guide to set up the end-to-end real-time fraud detection demo which is based on the IEEE-CIS financial open dataset. 

This solution is an open source project in GitHub, you can customize and optimize the model for your specific use cases.

Q: How much does it cost?

A: This solution is offered for free. You will be charged based on resources used in the cloud. 

Q: Why use graph database in this solution?

A: We use graph database to store the relationships between entities. The graph database provides the microseconds query performance to query the sub-graph of entities for real-time fraud detection inference.

Q: What’s the benefits of using the graph neural network in the scenario of Fraud Detection?

A: In the scenario of fraud detection, fraudsters can work as groups to hide their abnormal features but leave some traces of relations. Traditional machine leaning models use various features of samples. However, the relations among different samples are always ignored, either because of no direct feature can represent these relations, or the unique values of a feature is too big to be encoded for models. For example, IP addresses and physical addresses can be a link of two accounts. But the unique values of these addresses are too big to be one-hot encoded. Many feature-based models, hence, cannot leverage these potential relations.
Graph Neural Network models directly benefit from links built among different samples, once reconstruct some categorical features of a sample into different nodes in a graph structure. Via using message pass and aggregation mechanism, GNN-based models can not only use features of samples but also capture the relations among samples. With the advantages of capture relations, Graph Neural Network is more capable of detecting collaborated fraud event compared to traditional models.