Incremental Fraud Detection Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fraudulent transaction detection systems face challenges due to concept drift and verification latency, as existing machine learning algorithms require labeled historical data and fail to adapt to changing transaction patterns and customer behaviors over time.
Innovation Solution
A system and method for incrementally training classifiers using oracle prediction paradigms, where multiple classifiers are applied to different data subsets, allowing for continuous updating and labeling of transactions as fraudulent, thereby addressing concept drift and improving detection performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised or unsupervised machine learning algorithms are trained on labeled historical data, then fraud detection performance is improved initially, but detection performance deteriorates over time due to concept drift and verification latency
Solution Approach 1:
The system uses unlabeled transactions to automatically train and update classifiers without requiring manual investigator verification. The classifiers serve themselves by finding patterns in unlabeled data, eliminating the bottleneck of human verification while maintaining detection performance over time.
Solution Approach 2:
The system performs preliminary classification of unlabeled transactions using multiple classifiers before investigator verification. This preliminary action creates candidate fraudulent transaction lists that investigators can verify more efficiently, reducing overall verification latency while maintaining high detection accuracy.
2Adaptability or versatility
If multiple classifiers are applied to incremental datasets with incremental training, then adaptability to concept drift is improved, but device complexity increases
Solution Approach 1:
The system segments the classifier training process into independent parallel classifiers, each trained on different subsets of incremental data. This segmentation allows the system to handle concept drift through multiple specialized classifiers rather than one complex monolithic classifier, managing complexity through division of labor.
Solution Approach 2:
The system merges results from multiple independent classifiers to produce final fraud detection decisions. By combining the strengths of different classifiers trained on different data subsets, the system achieves high adaptability to concept drift while keeping individual classifier complexity manageable.
Data Source
AI summary
A system and method for detecting fraudulent transactions of incremental transactions via incrementally training classifiers on a number of data subsets to obtain transaction label predictions. Recent information on fraudulent transactions is implemented to update the classifiers, thereby eliminating the problems of concept drift and verification latency. Some of the incremental transactions may be labeled as fraudulent according to the incrementally modified classifiers and thereby rejected.