AI Model Routing Without Inference Engines for Fast Fraud Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for selecting computer models in electronic payment systems are inefficient and fail to provide rapid and accurate routing, often relying on static thresholds or binary analyses, which is particularly problematic in time-sensitive transactions.
Innovation Solution
An intelligent request routing system using machine-learning models to identify suitable computer models for transaction verification, considering factors like computational power, memory capacity, and compatibility, and employing a routing machine-learning model to adapt to specific sub-models for improved performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional static threshold methods are used for model routing, then system complexity is reduced, but recall and fraud detection accuracy deteriorate
Solution Approach 1:
The patent introduces a routing machine-learning model as an intermediary component between the transaction input and the fraud-detection models. This routing model evaluates multiple characteristics (transaction amount, merchant category, device information, etc.) and dynamically selects the most appropriate fraud-detection model, thereby improving detection accuracy without requiring the entire system to be complex. The intermediary handles the complexity of model selection internally while presenting a simple interface externally.
Solution Approach 2:
The patent segments the fraud-detection system into multiple specialized models, each optimized for specific transaction types or fraud patterns. The routing model divides incoming transactions into different segments and directs them to appropriate specialized models. This segmentation allows each model to focus on specific patterns, improving overall recall and accuracy while maintaining manageable system complexity through modular architecture.
2Reliability
If multiple fraud-detection models are executed for all transactions, then recall improves, but processing time and computational resources increase
Solution Approach 1:
The patent implements dynamic model routing where the selection of fraud-detection models is not fixed but adapts based on transaction characteristics. The routing model dynamically determines which models to execute for each transaction based on real-time evaluation of transaction amount, merchant category, device information, and other features. This dynamic approach ensures that only necessary models are executed, reducing processing time while maintaining high recall by selecting the most relevant models for each transaction type.
Solution Approach 2:
The patent changes the parameters of model execution by using a routing model that evaluates multiple transaction parameters (amount, merchant category, device info, location) to determine model selection. Instead of executing all models with fixed parameters for every transaction, the system adjusts which models are activated based on the specific parameter values of each transaction, thereby reducing unnecessary computational overhead while maintaining detection effectiveness.
3Adaptability or versatility
If a single fraud-detection model is used for all transactions, then system complexity is minimized, but adaptability to different transaction types deteriorates
Solution Approach 1:
The patent creates a universal routing machine-learning model that handles multiple transaction types and fraud patterns through a single multi-functional system. This routing model serves as a universal entry point that evaluates all incoming transactions and directs them to appropriate specialized fraud-detection models. The universality of the routing model allows the system to adapt to different transaction types without requiring separate routing logic for each model, thereby improving adaptability while controlling overall system complexity.
Data Source
AI summary
One method includes identifying a set of fraud-detection ML models to be used to determine a fraud status of a transaction; executing a routing ML model configured to generate a score for at least a subset of the set of fraud-detection ML models, based on an input of an amount and an attribute associated with the transaction, the score indicative of the fraud status for the transaction based on the amount and the attribute; executing using the amount and the attribute associated with the transaction, a subset of the set of fraud-detection ML models selected in accordance with each fraud-detection ML model's respective score, whereby at least one fraud-detection ML model within the subset of the set of fraud-detection models generates a prediction associated with the fraud status of the transaction; and authorizing the transaction based on the prediction.