Machine Learning Fraud Detection Using Relational Graph Convolutional Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods fail to effectively identify and prevent fraudulent transactions in both in-store and online retail settings, leading to financial losses for both customers and retailers due to the lack of real-time detection and verification processes.
Innovation Solution
The implementation of a machine learning-based system that generates a trust score for transactions by analyzing customer and device data using relational-graph convolutional networks, ensemble supervised learning processes, and customer segmentation models to determine the legitimacy of purchases, thereby preventing fraudulent transactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional transaction verification methods are used, then transaction processing is simple and fast, but fraudulent transactions cannot be effectively detected
Solution Approach 1:
The system segments fraud detection into multiple independent components: device graph analysis, customer graph analysis, transaction feature extraction, and machine learning model evaluation. Each component processes specific aspects of transaction data separately, then combines results to generate an overall fraud risk score, improving detection accuracy while maintaining manageable system complexity
Solution Approach 2:
The patent introduces intermediate data structures including device graphs that model relationships between devices, customers, and transactions. These graphs serve as intermediaries that transform raw transaction data into structured relationship information, enabling the machine learning models to detect fraudulent patterns that would be invisible in traditional verification systems
2Reliability
If real-time fraud detection is implemented, then fraudulent transactions can be prevented, but processing time and computational resources increase
Solution Approach 1:
The system performs preliminary actions by pre-computing device graphs and customer graphs during periods when transactions are not being processed. These graphs capture relationships between devices, customers, and transactions in advance. When a new transaction arrives, the system only needs to query these pre-computed graphs and apply fraud detection algorithms, rather than building relationships from scratch, significantly reducing real-time processing time
Solution Approach 2:
The patent applies partial action by implementing fraud detection with varying intensity based on risk assessment. For low-risk transactions, the system performs minimal verification using pre-computed graphs. For suspicious transactions identified through initial screening, the system applies more intensive analysis including additional graph queries and multiple machine learning model evaluations, optimizing the balance between detection accuracy and processing speed
Data Source
AI summary
This application relates to apparatus and methods for identifying fraudulent transactions. The embodiments may employ machine learning processes to detect fraudulent activity. In some examples, a computing device determines customer data and device data for a customer and device involved in a transaction. The customer data may include previous transactions by the customer, and the device data may include previous transactions involving the device. The computing device generates features based on the customer data and the device data, and applies one or more machine learning models to the generated features to generate a trust score. The trust score is indicative of how likely a transaction is to be fraudulent. In some examples, the transaction is not allowed if the trust score is beyond a threshold. In some examples, the computing device trains the machine learning models based on customer data and device data for a plurality of customers and devices.