HierTGAN: Hierarchical Time Series Generation with Aggregation Constraints

Abstract

Generative models for time series data have been able to preserve the temporal dynamics of the original time series and are extremely successful in generating realistic synthetic data. However, in the real world, time series data can be disaggregated by various attributes of interest, thereby forming a hierarchical structure, often referred to as hierarchical time series data. Existing models for time series generation do not capture the structural dynamics (inter-level relationships of the hierarchy) of hierarchical time series data. Therefore, in this research, for the first time, we introduce HierTGAN, an auto-regressive generative adversarial network (GAN) for hierarchical time series generation. The proposed HierTGAN solves for an equivalent inter-level relationship within the embedding space generated by an autoencoder. Multiple experiments have been performed to evaluate the effectiveness of HierTGAN in generating realistic synthetic hierarchical time series data. © 2024 ACM.