Zero-Inflated Time Series Generation

Zero-inflated time series (ZITS), characterised by excessive zeros relative to their standard distributions, pose significant challenges for synthetic data generation due to the need to simultaneously preserve the temporal dependencies, the excess zero structure, and the heavy-tailed non-zero distributions. Current generative models, including variational autoencoders (VAE) and generative adversarial networks (GAN), lack the mechanisms to handle zero-inflated data and fail to generate data with similar characteristics. This work proposes two architectures, ZITS-GAN and ZITS-VAE, with a two-head decoder that separately models the zero pattern and non-zero magnitude. The architectures use dilated convolutions; ZITS-VAE uses a β-weighted ELBO and ZITS-GAN uses a Wasserstein objective with gradient penalty. We evaluated the proposed models on three real-world datasets: monthly car parts sales, M5 retail sales, and household IoT device durations, with zero ratios of 76%, 59%, and 74% respectively. The proposed models reproduced zero ratios and autocorrelation more closely than several deep generative baselines, while results on other distributional metrics varied by dataset and model.