Fraud Detection via Anomaly Score Interaction and PCA
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fraud detection systems rely on unsupervised models to identify anomalies, but lack a benchmark for anomalies, leading to difficulties in distinguishing real anomalies from false positives, and ensemble methods often miss anomalies on the edge of model thresholds.
Innovation Solution
The approach generates anomaly score variables using multiple unsupervised models, normalizes these scores, constructs interactions between them, and applies principal component analysis (PCA) to detect anomalies, providing a more comprehensive and accurate identification of fraudulent transactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If unsupervised models are used to detect fraudulent transactions without known examples, then the system can identify potential anomalies, but the system lacks a benchmark for anomalies leading to difficulties in distinguishing real anomalies from false positives
Solution Approach 1:
The patent combines multiple unsupervised models (Isolation Forest, Local Outlier Factor, One-Class SVM) to generate ensemble anomaly scores. By merging the strengths of different algorithms, the system achieves more reliable anomaly detection with reduced false positives, as each model compensates for the weaknesses of others.
Solution Approach 2:
The patent transforms anomaly scores from multiple models into standardized z-scores through normalization. This parameter transformation allows for meaningful comparison and combination of scores from different models with different scales, improving the precision of anomaly identification by establishing a common benchmark.
2Reliability
If multiple unsupervised models are used to generate anomaly scores, then the detection becomes more comprehensive, but the complexity of the system increases
Solution Approach 1:
The patent applies z-score normalization to transform anomaly scores from multiple complex models into standardized parameters with mean=0 and standard deviation=1. This parameter transformation simplifies the aggregation process and makes the system more manageable while preserving the comprehensive detection capability.
Solution Approach 2:
The patent introduces an intermediary normalization layer between the multiple unsupervised models and the final anomaly detection. This intermediary component (z-score calculation) mediates the complexity by providing a standardized interface for combining results from diverse models, reducing overall system complexity.
3Measurement precision
If anomaly scores are normalized and interactions are constructed, then the identification of real anomalies improves, but the computational requirements and processing time increase
Solution Approach 1:
The patent calculates interactions between normalized anomaly scores but applies thresholding to focus only on significant interactions. By performing partial action (calculating only necessary interactions above certain thresholds), the system maintains high precision in anomaly identification while reducing unnecessary computational overhead and processing time.
Data Source
AI summary
An approach is provided in which the approach generates anomaly score variables using multiple unsupervised models based on a set of data records. The approach normalizes the anomaly score variables into multiple normalized variables, and constructs at least one interaction based on a first one of the normalized variables and a second one of the normalized variables. The first normalized variable corresponds to a first one of the anomaly score variables and the second normalized variable corresponds to a second one of the anomaly score variables. The approach detects a set of anomalies based on the at least one interaction and transmits the set of anomalies to a user.