In recent years, the tabular generative model has emerged as a major research area across finance applications (e.g., fraud detection), especially under circumstances where original data is unavailable due to financial regulations such as the General Data Protection Regulation (GDPR). However, there is a notable scarcity of research on tabular generative models that directly handle the three characteristics of finance fraud detection (FFD) data toward more effective financial fraud detection: (a) extreme class imbalance, (b) high data sparsity, and (c) a large number of attributes with non-normal distribution. Through a preliminary study, we confirmed that the performance of existing tabular generative models declines as the intensity of the three characteristics increases. To address these challenges, we propose novel graph-theoretical tabular generative models for tabular data generation, which are specialized for financial fraud detection. More specifically, we implement three variants of our method— SeparateGGM , SignedGGM , and HybridGGM —which are tailored to capture class-wise structure, signed inter-class relationships, and their integration, respectively. Our models jointly leverage several graph-theoretical methodologies, including graph neural network-based feature augmentation, graph topology and connectivity analysis, and new graph centrality indicators that we design in this work. Especially, through the graph analysis, we empirically find that certain hyperparameter values (i.e., positive and negative K ratios in separate and signed K-NN graphs), which make the related graph measurements approximately the highest points, consistently yield the highest detection accuracy, thus facilitating the selection of effective graph settings. Experimental results show that our method outperforms seven different competing methods across three popular FFD benchmarks and one private dataset. Finally, we plan to deploy SignedGGM , the best efficient variant of our method, and HybridGGM-A , the best performing variant of our method, in a real-world financial application in the Financial Security Institute (FSI), one of South Korea's financial government agencies.
Park et al. (Thu,) studied this question.