Evaluating Time Series Classification with GAN-Generated Synthetic Data

In supervised machine learning, accurate classification tasks depend on the availability of diverse and extensive datasets. Nonetheless, data scarcity compromises the consistency and reliability of these tasks. This study addresses this challenge by evaluating Generative Adversarial Networks (GANs) for augmenting time series data when data is limited. We employed three GAN architectures, RCGAN, SIGCWGAN, and RTSGAN, to generate synthetic samples from ten unidimensional, balanced time series datasets. After integrating these samples into the original datasets, we evaluated the classification performance using five different classifiers, measuring accuracy and F1 score. Our findings reveal that GANs’ effectiveness in improving time series classification varies by dataset and architecture. Some GANs marginally enhanced accuracy and F1 scores, but this was not consistent across all models, with occasional performance declines.