Simplified Fraud Models From Complex Outputs for Real-Time Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex machine learning models for fraud detection are cumbersome to update and require significant computational resources due to their complexity, comprising multiple sub-models of different types, making them inefficient for real-time fraud prediction.
Innovation Solution
A simpler machine learning model is trained using the predictions of a complex model, comprising a single sub-model, and validated to ensure accuracy by comparing outputs with the complex model, allowing for efficient updates and reduced resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a complex machine learning model with multiple sub-models is used to accurately detect subtle fraud indicators, then fraud detection accuracy is improved, but device complexity and computational resource requirements increase
Solution Approach 1:
The patent creates simplified copies of the complex fraud detection model by training multiple simple models on different subsets of features. These simple model copies collectively replicate the detection capabilities of the original complex model while requiring fewer computational resources individually. The copying principle is applied by generating ensemble predictions from multiple simplified model instances rather than using a single complex model.
Solution Approach 2:
The patent segments the feature space by dividing all features into multiple subsets, each processed by a different simple machine learning model. This segmentation allows the complex detection task to be broken down into simpler sub-tasks that can be handled by less complex models. The segmentation of features enables parallel processing while maintaining comprehensive fraud detection coverage.
2Reliability
If a complex machine learning model is used to detect subtle fraud indicators, then fraud detection accuracy is improved, but updating and tuning processes become more complex and time-consuming
Solution Approach 1:
The model updating process is segmented by allowing individual simple models to be updated independently based on their specific feature subsets. This segmentation enables targeted updates without requiring retraining of the entire complex model system. Each simple model can be tuned and updated separately, significantly reducing the complexity and time required for model maintenance while preserving overall detection accuracy.
3Measurement precision
If a complex machine learning model is used to generate fraud predictions, then prediction accuracy is improved, but computational resources and processing time increase
Solution Approach 1:
The computational workload is segmented across multiple simple models that process different feature subsets in parallel. This segmentation distributes the computational burden, allowing predictions to be generated more efficiently than a single complex model would require. The parallel processing of segmented features reduces overall computational resource usage while maintaining prediction accuracy through ensemble aggregation.
Solution Approach 2:
Instead of running a single complex model that consumes substantial computational resources, the patent creates multiple simplified model copies that collectively provide the same detection capability. These simpler model copies require fewer resources each, and their combined operation achieves comparable accuracy to the complex model while being more computationally efficient overall.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a computer implemented method for training a simple machine learning model 210 to predict fraud. The method comprises obtaining a data set 120 comprising a plurality of data records 125, inputting the data set 120 into a complex machine learning model 110 to generate a fraud prediction score 140 for each of the data records 125, labelling the data records 125 as being indicative of fraudulent activity 150 or not being indicative of fraudulent activity 160 based on the fraud prediction score 140 generated for each respective data record 125, and training the simple machine learning model 210 on the labelled 150, 160 data records 125. The trained simple machine learning model 210 can then be used to predict fraudulent activity and determine if an alert 355 of the fraudulent activity is required The complex machine learning model 110 comprises a plurality of sub-models 113, 115, and the complexity of the complex machine learning model is due to this plurality of sub-models 113, 115. Whereas the simple machine learning model 210 comprises a single sub-model and is therefore substantially less complex than the complex machine learning model 110.