Reinforcement Learning for Optimizing Conditional Tabular Generative Adversarial

Financial fraud datasets typically exhibit a pro-nounced class imbalance, leading to elevated false positive rates when employing machine learning classifiers. To mitigate this issue, the Conditional Tabular Generative Adversarial Network (CTGAN), an oversampling technique tailored for tabular data, has emerged as a promising solution by generating additional minority samples. However, it remains pertinent to investigate whether optimizing the hyperparameters governing CTGAN can yield further improvements in performance. In light of the recent application of reinforcement learning (RL) for hyperparameter tuning, this study delves into the potential of harnessing RL mod-els to fine-tune the hyperparameters of CTGAN. The overarching objective is to design an RL-based methodology that automates the optimization of CTGAN hyperparameters. Although the initial findings from this endeavor may be inconclusive, they pave the way for future research avenues, exploring the potential of RL in the context of CTGAN hyperparameter tuning and scrutinizing the nuanced effects of CTGAN hyperparameters on classification performance.