Fraud Risk Self-Labelling Using Graph-Based Account Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in effectively identifying and mitigating financial fraud using machine learning models, particularly in assigning accurate training labels to accounts based on historical data and transaction patterns.
Innovation Solution
A method and system for assigning machine learning training labels to accounts by identifying suspicious behavior through account history data matching, PII similarity, and financial transaction activity, utilizing graph algorithms to cluster accounts and determine risk levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used for fraud detection, then fraud detection capability is improved, but accuracy in assigning training labels deteriorates due to lack of ground truth data
Solution Approach 1:
The system performs preliminary actions by identifying suspicious accounts through pattern matching before actual fraud occurs. It proactively clusters accounts with similar PII and transaction patterns, and pre-labels them as fraudulent or victim based on their relationships to confirmed fraud cases, enabling future ML models to learn from these pre-prepared labels.
Solution Approach 2:
The system introduces an intermediary labeling mechanism that uses graph-based account clustering and PII similarity analysis as a bridge between raw transaction data and ML training labels. This intermediary process infers fraud status for unlabeled accounts by their relationships to confirmed fraud cases, creating accurate training labels without direct ground truth.
2Measurement precision
If graph algorithms are used to cluster accounts by PII similarity, then fraud network identification is improved, but computational complexity increases
Solution Approach 1:
The system segments the account population into distinct clusters based on PII similarity using graph algorithms. By dividing accounts into groups with shared personal information, it identifies potential fraud networks more efficiently than analyzing all accounts individually, reducing the computational burden while maintaining identification accuracy.
Solution Approach 2:
The system changes parameters by using PII similarity as a clustering criterion instead of traditional transaction-based grouping. By transforming the problem into a graph-based similarity analysis with adjustable similarity thresholds, it achieves better fraud network identification while managing computational complexity through parameter optimization.
3Measurement precision
If multiple filtering stages are applied to identify suspicious accounts, then fraud detection accuracy is improved, but processing time increases
Solution Approach 1:
The system performs preliminary filtering by identifying accounts that match known fraud patterns before applying more complex analysis. This preliminary stage quickly eliminates obvious non-fraud cases, allowing subsequent detailed analysis to focus only on suspicious accounts, thereby maintaining high accuracy while reducing overall processing time.
Solution Approach 2:
The system applies partial filtering by focusing computational resources on accounts with the highest suspicion scores after initial pattern matching. Instead of applying all filtering stages uniformly to all accounts, it selectively intensifies analysis only where needed, achieving high detection accuracy without proportionally increasing processing time for the entire account population.
Data Source
AI summary
The presently disclosed subject matter relates to detection and mitigation of financial fraud, and in particular to implementation of systems for training machine learning models in such systems. It involves self-labelling of fraud risk in a transaction processing system.