UniTSGAN: A Unified Transformer-based Framework for Imbalanced Time Series Generation and Classification

Junzhi Wen, Rafal Angryk

Keywords: time series classification, synthetic data generation, class imbalance, rare events, transformers, GANs

Abstract: Handling severe class imbalance in time series data is a critical challenge, especially for rare-event prediction in domains such as space weather forecasting and health monitoring. Standard discriminative models often perform poorly on the underrepresented minority class, which typically represents the most important outcomes for decision-making. Common remedies like oversampling or undersampling can cause overfitting or information loss. Generative Adversarial Networks (GANs) have shown promise in generating realistic synthetic data, but they are generally not optimized for class-discriminative generation and must be combined with separate classifiers. In this paper, we propose UniTSGAN, a unified adversarial framework that jointly handles multivariate time series generation and binary classification in highly imbalanced scenarios. Our model uses a transformer encoder for both the generator and discriminator. The generator can be pretrained with an unsupervised masking-based objective to learn latent representations of the minority class, and the discriminator has a dual-head architecture that simultaneously performs real-vs-fake discrimination and class label prediction. This design allows the model to learn realistic, class-consistent synthetic samples and a robust classifier in a single training process. To evaluate generative performance, we introduce a classification-based metric that measures how much adding synthetic data improves downstream classification. Experimental results on seven real-world datasets demonstrate that UniTSGAN consistently outperforms state-of-the-art methods on both imbalanced time series classification and generation tasks, particularly in low-data regimes.

Primary Area: learning on time series and dynamical systems

Submission Number: 23604