BalancerGNN: Balancer Graph Neural Networks for imbalanced datasets: A case study on fraud detection

Area of Science:

• Machine Learning
• Data Science
• Graph Neural Networks

Background:

• Fraud detection on imbalanced datasets is challenging due to model bias towards majority classes.
• Data imbalance negatively impacts graph construction, a critical step for Graph Neural Networks (GNNs).

Purpose of the Study:

• To introduce the BalancerGNN framework to effectively handle imbalanced datasets in fraud detection.
• To demonstrate the framework's superiority over existing methods in identifying fraudulent cases.

Main Methods:

• Developed a three-component framework: node construction (Graph-based Variable Clustering and Encoder-Decoder based Dimensionality Reduction), balanced neighbor sampling for graph construction, and GNN training with balanced batches and a custom loss function.
• Utilized transformer-based techniques for feature representation and dimensionality reduction.
• Employed balanced training batches and a multi-component loss function for GNN training.

Main Results:

• BalancerGNN achieved high sensitivity rates (72.87%–81.23%) and accuracy (73.99%–94.28%) across Medicare, Equifax, IEEE, and auto insurance fraud datasets.
• The framework consistently outperformed other methods in identifying fraud cases.
• Node construction and balanced neighbor sampling were highlighted as crucial for performance.

Conclusions:

• BalancerGNN effectively addresses the challenges of fraud detection in imbalanced datasets.
• The proposed methods for node construction, graph representation, and neighbor sampling significantly enhance GNN performance for fraud detection.
• The framework shows strong potential for real-world fraud detection applications.