Knowledge Graph Fraud Detection via GNN Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fraud detection systems lack the ability to analyze data across different modalities, limiting the types of patterns they can detect and the actions they can take, especially when data is coming in from multiple channels.
Innovation Solution
A method and system using a knowledge graph and graph neural networks to detect fraudulent patterns in financial transactions across multiple data modalities, allowing for automatic fraud limiting actions to be taken.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If fraud detection systems use preconfigured filters to analyze payment records, then they can identify potential fraud cases using simple patterns, but they lack the ability to analyze data across different modalities and detect complex fraudulent patterns
Solution Approach 1:
The patent merges multiple data modalities (transactions, devices, locations, timestamps) into a unified knowledge graph structure where nodes represent entities and edges represent relationships. This integration allows the system to analyze cross-modal patterns that would be invisible to traditional single-modality filter-based systems, directly resolving the contradiction between versatility and complexity.
Solution Approach 2:
The patent transforms flat transaction data into a multi-dimensional knowledge graph embedding space where fraudulent patterns emerge as geometric structures (clusters, anomalies, subgraphs). This dimensional transformation enables detection of complex fraud patterns across multiple modalities simultaneously, converting the versatility-complexity tradeoff into a geometric pattern recognition problem.
2Reliability
If fraud detection systems analyze multiple data channels (web data, call data, etc.), then they can detect more comprehensive patterns, but they are limited in the types of actions they can take
Solution Approach 1:
The patent implements dynamic fraud response where the system automatically adjusts its actions based on the detected pattern type and confidence level. The knowledge graph embedding analysis enables the system to dynamically select from multiple response actions (block transaction, request verification, alert user, investigate further) rather than being limited to fixed preconfigured responses, thus increasing both reliability and adaptability.
Solution Approach 2:
The system incorporates feedback loops where detection results inform subsequent actions and system configuration. The automated action-taking based on detected patterns creates a closed-loop system that continuously improves detection accuracy and response effectiveness, allowing the system to adapt its behavior based on incoming data from multiple channels.
3Measurement precision
If fraud detection systems use traditional analysis methods, then they are computationally efficient, but they cannot detect sophisticated fraudulent patterns across multiple data modalities
Solution Approach 1:
The patent performs preliminary embedding of the entire knowledge graph into a compressed vector space representation before pattern detection. This preprocessing step transforms the complex multi-modal data into a compact embedding format that captures structural relationships, enabling efficient subsequent pattern search and analysis. The embedding computation is performed once and can be reused for multiple detection queries, reducing overall computational resources while maintaining high detection precision.
Data Source
AI summary
A system and method for tasks assistance using a learning module comprising a knowledge graph and associated graph neural network is disclosed. The system can represent data using a knowledge graph, and generate embeddings of the knowledge graph for detecting latent patterns in the data that may be obscured in a high dimensional representation of the data. The system can more readily detect fraud patterns and take appropriate fraud limiting actions.