SFA-GNN Fraud Detection via Neighbor Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection models, particularly graph neural network (GNN) based systems, face challenges in accurately identifying fraudulent transactions due to class imbalance issues, where a small fraction of fraudulent transactions are masked among numerous non-fraudulent ones, leading to poor performance in heavily imbalanced datasets.
Innovation Solution
The proposed solution involves a Split-Filter-Aggregate Graph Neural Network (SFA-GNN) model that splits a node's neighborhood into label-aware sub-graphs, uses a Siamese Neural Network (SNN) for filtering neighbors based on contrastive scores, and aggregates information from consistent neighbors to improve the accuracy of fraudulent transaction detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If GNN aggregation assigns equal importance to all neighbors, then the model is simple to implement, but fraudsters can easily camouflage among benign entities and evade detection
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different neighbor nodes based on their fraud risk characteristics. Instead of uniform aggregation, the model assigns different weights to different neighbors, giving higher importance to suspicious nodes and lower importance to benign nodes. This localized differentiation enables the model to detect fraudsters who camouflage among benign entities while maintaining the overall GNN framework.
2Loss of information
If the model processes all neighbor nodes in GNN aggregation, then complete information is captured, but computational complexity increases significantly
Solution Approach 1:
The patent extracts and identifies the most critical subset of neighbor nodes that contribute most to fraud detection performance. By using the fraud risk score to select only the top-k suspicious neighbors for aggregation, the model removes unnecessary computational overhead from processing all benign neighbors while retaining the essential information needed for accurate fraud detection.
3Quantity of substance
If re-sampling and re-weighting techniques are used to balance class distribution, then training data balance improves, but these methods are not effective for GNN-based models with heavy class-imbalance
Solution Approach 1:
The patent changes the parameter of neighbor selection from uniform sampling to risk-based sampling. Instead of treating all neighbors equally or using traditional re-sampling methods, the model dynamically adjusts the sampling strategy based on the fraud risk scores calculated for each neighbor. This parameter change enables effective handling of class imbalance in GNN models by focusing computational resources on the most relevant suspicious nodes.
Data Source
AI summary
Methods and server systems for detecting fraudulent transactions are described herein. Method performed by server system includes accessing base graph including plurality of nodes further including plurality of labeled nodes and unlabeled nodes. Method includes assigning via Graph Neural Network (GNN) model, fraudulent label or non-fraudulent label to each unlabeled node based on the base graph. This assigning process includes generating plurality of sub-graphs based on splitting the base graph and filtering these sub-graphs via Siamese Neural Network model based on pre-defined threshold values. Then, the GNN model generates plurality of sets of embeddings based on plurality of filtered sub-graphs. Further, aggregated node embedding is generated for each node and then, final node representation for each node is generated via dense layer of GNN model. Then, fraudulent label or the non-fraudulent label is assigned to each unlabeled node of plurality of unlabeled nodes based on final node representation.