Abstract
Time series generation models are crucial for applications like dataaugmentation and privacy preservation. Most existing time series generationmodels are typically designed to generate data from one specified domain. Whileleveraging data from other domain for better generalization is proved to workin other application areas, this approach remains challenging for time seriesmodeling due to the large divergence in patterns among different real worldtime series categories. In this paper, we propose a multi-domain time seriesdiffusion model with domain prompts, named TimeDP. In TimeDP, we utilize a timeseries semantic prototype module which defines time series prototypes torepresent time series basis, each prototype vector serving as "word"representing some elementary time series feature. A prototype assignment moduleis applied to extract the extract domain specific prototype weights, forlearning domain prompts as generation condition. During sampling, we extract"domain prompt" with few-shot samples from the target domain and use the domainprompts as condition to generate time series samples. Experiments demonstratethat our method outperforms baselines to provide the state-of-the-art in-domaingeneration quality and strong unseen domain generation capability.