Fraud Detection Model Selection via Segmented Transaction Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fraud detection systems in payment transactions face challenges in adapting to dynamic fraud patterns, particularly in card-not-present transactions, as they often rely on a single modeling technology that is not effective over time and can result in false positives, denying legitimate transactions.
Innovation Solution
A fraud detection system that rapidly trains and selects multiple predictive modeling technologies to best address the current fraud environment, automatically choosing the most effective model based on the fraud environment and available resources, allowing for daily or weekly updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single modeling technology is used for fraud detection, then the system is simple to operate and maintain, but it cannot effectively adapt to dynamic fraud patterns over time
Solution Approach 1:
The patent segments the fraud detection task by dividing it into multiple independent modeling technologies (e.g., rule-based models, machine learning models, statistical models). Each model type handles specific fraud patterns independently, allowing the system to adapt to different fraud scenarios without requiring a complete system overhaul. This segmentation enables flexible adaptation while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent implements a universal fraud detection framework that incorporates multiple modeling technologies capable of handling diverse fraud patterns. The system is designed to be multi-functional, where different model types can be applied to various transaction scenarios (e-commerce, mobile commerce, card-not-present transactions). This universality allows a single system to address multiple fraud types without requiring separate specialized systems for each.
2Reliability
If fraud detection algorithms are tightened to detect more fraud, then fraud detection capability improves, but legitimate transactions are denied and customer experience deteriorates
Solution Approach 1:
The patent applies local quality by using different modeling technologies with different sensitivity levels for different fraud scenarios and transaction types. Rather than applying a uniform high-threshold detection algorithm to all transactions, the system selectively applies rigorous detection only where fraud risk is highest, while allowing smoother processing for low-risk transactions. This localized approach maintains high fraud detection accuracy while preserving customer experience for legitimate users.
Solution Approach 2:
The patent dynamically adjusts detection parameters and thresholds based on the specific fraud pattern being detected and the transaction context. The system can modify sensitivity parameters, confidence thresholds, and decision criteria to balance fraud detection and customer experience. This parameter flexibility allows the system to tighten detection when fraud is suspected while loosening it for legitimate transactions, thereby maintaining both reliability and ease of operation.
3Reliability
If multiple modeling technologies are trained and selected automatically, then fraud detection effectiveness improves, but the computing resources and training time required increase
Solution Approach 1:
The patent performs preliminary action by pre-training multiple modeling technologies offline before deployment. The system prepares a library of pre-trained models covering various fraud patterns, so that during actual transaction processing, the system only needs to select and apply the appropriate pre-trained model rather than training from scratch. This preliminary training approach shifts the computational burden to offline periods, reducing real-time resource consumption while maintaining high detection effectiveness.
Solution Approach 2:
The patent uses copying by creating multiple instances of modeling technologies with different configurations and training data. Instead of developing entirely new models for each fraud scenario, the system creates copies or variants of base models that can be quickly adapted to specific fraud patterns. This copying approach reduces the overall computational resources needed, as the foundational model architecture and training pipelines can be reused across multiple model instances.
Data Source
AI summary
A payment processing system continuously processes transactions. Multiple modeling technologies are used within a fraud detection system to each handle different blocks of segmented transaction data in order to detect fraud. Model data is created from recent segmented transaction data that is currently handled by an existing model; each transaction includes a fraud flag as to whether it was fraudulent or not. The model data is used to train, validate and test any of a number of types of models where the fraud flag is the target. Performance metrics such as Sensitivity, Review Rate and False Positives are calculated for each trained model. An operational objective is used to determine a value for a particular performance metric which then dictates values for other performance metrics for each model. The best model is chosen by optimizing the performance metrics. The best model is configured and deployed to replace the existing model.