TRGAN: A Time-Dependent Generative Adversarial Network for Synthetic Transactional Data Generation
Kirill Mikhailovich Zakharov, Elizaveta Stavinova, Anton Lysenko · 2023
Synthetic transactional data generation methods are in high demand nowadays as they enable organizations to extract knowledge from data, develop robust data-driven models, and make informed decisions without compromising sensitive information present in real-world data. In this paper, we propose TRGAN, a method for generating synthetic bank transactions. Our method features a robust GAN-based architecture with a Supervisor component that controls the training process and utilizes a conditional vector for precise attribute generation within specific time intervals. Additionally, our approach effectively captures global temporal attribute relationships by considering all attributes, including the time factor. Our experiments show that (1) synthetic transactional data can be effectively generated by the proposed method, (2) the time intervals between transactions generated by the proposed method are more accurate in comparison with Banksformer, (3) the proposed method outperforms Banksformer, CTGAN, and CopulaGAN across various evaluation metrics for categorical and numerical attributes.