CTCN: a novel credit card fraud detection method based on Conditional Tabular Generative Adversarial Networks and...

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Credit card fraud poses significant financial risks to individuals and institutions.
• Existing fraud detection methods struggle with imbalanced datasets and complex transaction patterns.

Purpose of the Study:

• To propose and evaluate a novel credit card fraud detection method, CTCN.
• To enhance fraud detection accuracy by addressing data imbalance and temporal dependencies.

Main Methods:

• Utilized Conditional Tabular Generative Adversarial Networks (CTGAN) for oversampling and data balancing.
• Implemented Neighborhood Cleaning Rule (NCL) to refine the dataset by removing overlapping majority class samples.
• Employed Temporal Convolutional Networks (TCN) to analyze transaction sequences and capture long-term dependencies.

Main Results:

• CTCN effectively generated synthetic minority class samples, creating a balanced dataset.
• TCN successfully identified relationships within transaction sequences for improved fraud detection.
• Experimental results on three public datasets showed CTCN outperformed existing machine and deep learning methods.

Conclusions:

• The proposed CTCN method offers a robust solution for credit card fraud detection.
• CTCN demonstrates superior performance in terms of recall, F1-Score, and AUC-ROC compared to current state-of-the-art techniques.