Continuous Learning Neural Network Using Rolling Window for Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection systems in distributed commercial transactions face challenges in adapting to concept drift and reacting quickly to dynamic network attacks, as they often rely on slow data processing and are reactive rather than proactive.
Innovation Solution
The system employs a neural network model that continuously learns and rebuilds by generating pseudo interaction data based on historical trends, embedding current and pseudo interaction data to form interaction data matrices, allowing for real-time adaptation and improved predictive power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional fraud detection systems process all historical data to train models, then model accuracy is improved, but system response time to new threats increases and adaptability to concept drift deteriorates
Solution Approach 1:
The patent segments historical data into concept drift intervals using a rolling window approach. Data is divided into recent window data and historical reference data, allowing the system to process only relevant recent data for training while maintaining awareness of historical patterns through concept drift detection. This segmentation enables fast response to new threats without being overwhelmed by all historical data.
Solution Approach 2:
The system dynamically adjusts the training data composition based on concept drift detection. When concept drift is detected, the system switches from using all historical data to using a rolling window of recent data combined with generated pseudo-interaction data. This dynamic adaptation allows the system to maintain high accuracy while responding quickly to evolving threat patterns.
2Loss of time
If the system uses a rolling window of recent data for training, then response time to new threats is improved, but the amount of training data decreases potentially affecting model accuracy
Solution Approach 1:
The patent generates pseudo-interaction data by copying and transforming historical reference data into the format of recent window data. This synthetic data copying process enriches the limited recent window data with patterns from historical data, effectively increasing the training data volume without compromising the system's ability to respond quickly to new threats. The pseudo-interaction data mirrors real interaction patterns while adapting to current concept drift.
3Adaptability or versatility
If the system continuously retrain models with new data, then adaptability to concept drift is improved, but computational resources and processing time increase
Solution Approach 1:
The system performs preliminary action by generating pseudo-interaction data from historical reference data before actual fraud detection operations. This pre-generated pseudo data is then used during training, reducing the computational burden of processing all historical data in real-time. The preliminary generation of training data from historical patterns allows for efficient continuous learning without excessive computational resource consumption.
4Measurement precision
If the system processes all historical interaction data, then comprehensive pattern recognition is improved, but the system becomes reactive rather than proactive in detecting new attack patterns
Solution Approach 1:
The system dynamically shifts its data processing focus based on the operational context. Instead of statically processing all historical data, the system dynamically selects between using recent window data for proactive detection and historical reference data for pattern generation. This dynamic approach enables the system to be proactive in detecting new attack patterns while maintaining comprehensive pattern recognition through the generated pseudo-interaction data.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A disclosed method an analysis computer determining a rolling window associated with interaction data for interactions that occur over time. The analysis computer can retrieve interaction data for interactions occurring in the rolling window. The analysis computer can then generate pseudo interaction data based upon historical interaction data. The analysis computer can optionally embed the interaction data for the interactions occurring within the rolling window and the pseudo interaction data to form interaction data matrices. The analysis computer can then form a neural network model using the interaction data matrices, which is derived from the interaction data in the rolling window and the pseudo interaction data.