Real-Time Fraud Model Updating with Validated User Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection models become outdated quickly and require significant computing resources to update, leading to delayed recognition of fraudulent transactions and false claims, which consume network and computing resources.
Innovation Solution
A system that enables real-time updating of a fraud model by sending immediate transaction notifications to users for confirmation, collecting user feedback, and validating the response to improve the model's accuracy using machine learning algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-time user feedback is collected for every transaction, then model accuracy improves, but network resources and user device resources are consumed
Solution Approach 1:
The system changes the parameter of notification timing from batch processing to real-time triggering based on fraud score thresholds. When a transaction exceeds a predetermined fraud threshold, immediate notification is sent to the user device, converting continuous monitoring into event-driven updates that reduce overall resource consumption while maintaining detection accuracy.
Solution Approach 2:
The system enables user devices to autonomously determine transaction validity by comparing received transaction data against their own records and context. The user device independently validates whether a transaction is legitimate without requiring centralized verification, reducing network communication overhead and server processing resources.
2Measurement precision
If fraud model is updated frequently, then detection accuracy improves, but computing resources are consumed
Solution Approach 1:
The system implements periodic model updates triggered by accumulated user feedback rather than continuous retraining. Feedback from multiple transactions is aggregated and used to update the fraud model at scheduled intervals, converting continuous computationally intensive operations into periodic updates that maintain accuracy while conserving computing resources.
Solution Approach 2:
The system pre-calculates fraud scores using the existing model before user feedback arrives, and prepares notification data in advance. This preliminary processing allows the system to respond quickly when feedback is received without requiring intensive real-time computation during the actual update process.
3Measurement precision
If user notifications are sent for all transactions, then feedback quality improves, but user experience deteriorates
Solution Approach 1:
The system changes the notification trigger parameter from universal to conditional based on fraud risk levels. Notifications are only sent when transactions exceed predetermined fraud thresholds, filtering out low-risk transactions that would unnecessarily interrupt users while maintaining high-quality feedback collection for suspicious activities.
Solution Approach 2:
The system uses disposable, context-specific notification messages that are discarded after serving their purpose. Each notification contains only the essential transaction information needed for validation, avoiding persistent or repetitive alerts that would burden users, while still collecting the necessary feedback data.
Data Source
AI summary
In some implementations, a device may receive exchange information associated with an exchange initiated by a medium that indicates a first location associated with the exchange and may enable the exchange to be completed. The device may transmit, to the user device, an indication of the exchange and a request to indicate whether the exchange is fraudulent. The device may receive, from the user device, a response indicating whether the exchange is fraudulent that indicates a second location associated with the user device. The device may determine whether the response is valid based on a comparison between the first location and the second location. The device may determine whether the exchange is fraudulent based on the response and determining whether the response is valid. The device may train the model using the exchange information and an indication of whether the exchange is fraudulent as an input to the model.