Tabular GAN-Based Oversampling of Imbalanced Time-to-Event Data for Survival Prediction

Class imbalance causes an underestimation (overestimation) of the hazard of minority class in survival prediction. A common strategy to handle class imbalance is to oversample the minority class by generating synthetic samples. This paper explores the potential of tabular Generative Adversarial Networks (GANs) for oversampling based on real world survival datasets and simulated imbalanced datasets. We compare GAN-based oversampling methods with traditional methods on generation of minority instances and balanced survival prediction. Experimental results show that balanced survival prediction after GAN-based oversampling can outperforms baseline in some situations, and also demonstrate that traditional oversampling methods perform better than GAN-based methods on both minority samples generation and balanced survival prediction.