Financial Data Generation Utilizing Graph Information from Transaction Networks
Kim Kwanggeun, Seonkyu Lim, Seungho Choi, Suk‐Ju Kang · 2025
High-quality data is crucial for the advancement of machine learning and deep learning models. However, in the financial domain, the amount of data available for training models is limited due to privacy concerns and financial regulations. To address these challenges, the financial industry has increasingly turned to synthetic data generation methods. Previous approaches to generating synthetic financial data, including GAN-based models, have focused on replicating the properties of tabular data. However, they do not perform well in preserving complex network structures, such as degree centrality, which are a significant information of financial transactions. To address this limitation, we propose a novel method for synthetic data generation using Augmented Financial Network (FNet), a self-developed module to preserve the degree centrality of the original data. By applying our FNet to existing GAN-based models, we generate synthetic data that accurately captures both the network structures and tabular properties of real-world financial transactions. Experimental results demonstrate that incorporating FNet into GAN-based models significantly improves fraud detection performance.