Generating time series data is important for many applications, including data augmentation, synthetic datasets, and scenarios. However, when there is more than one, this process becomes too complex because it includes variations of patterns across categories in the real world. With such wide variations in patterns among real-world categories, the complexity of the process tends to increase. The process especially becomes complicated because the data may not rely on historical records. It often falls apart in attempts to use natural language to describe domains when such descriptions are typically vague, incomplete, or impractical, especially for new or evolving areas.
Current time series generation methods employ models like GANs VAEs and mixed approaches like flows and ODEs. GANs are designed to optimize temporal dynamics, while VAEs focus on trend and seasonal decomposition using specialized decoders. Mixed methods attempt to unify diverse techniques but often fail to scale across multiple domains. Diffusion models like DDPMs generate data by reversing noise processes but mainly focus on single-domain settings. Multi-domain approaches rely on pretraining models on large datasets or normalizing data. However, they do not explicitly address the differences between domains and thus are less effective in handling diverse and evolving real-world challenges.
To tackle the challenge of generating time series from multiple domains while preserving the model’s ability to differentiate between them, researchers from Nanjing University, Microsoft Research Asia, and Peking University introduced a novel multi-domain time series diffusion model, TimeDP. This model utilizes time series semantic prototypes to define the basis of the time series, where each prototype vector represents an elementary time series feature. The model extracts domain-specific prototype weights by employing a prototype assignment module, which helps learn domain prompts as generation conditions. During the sampling process, domain prompts are generated using few-shot samples from the target domain. This ensures that the generated time series has the characteristics of the specific domain.
The researchers applied a training strategy that involved data from several domains. The strategy used conditional denoising and prototype assignment as the guiding process of generation. The model captured a diverse distribution of time series data by leveraging data from multiple domains. The model generated a time series for a selected domain by conditioning on domain-specific prototype assignments and using domain prompts. Furthermore, the approach supported generating time series from unseen domains by utilizing prototypes as a universal representation, enabling the model to generalize beyond the training data.
Researchers evaluated 12 datasets across four domains: Electricity, Solar, Wind (energy), Traffic, Taxi, Pedestrian (transport), Air Quality, Temperature, Rain (nature), and NN5, Fred-MD, Exchange (economic). The datasets were pre-processed into uni-variate sequences of 24, 96, 168, and 336. Using a multi-domain dataset, they compared their model with baselines like TimeGAN, GT-GAN, TimeVAE, and TimeVQVAE. Results showed that the proposed model outperformed others in generating time series closest to real data, with the best performance on MMD, K-L, and MDD. It surpassed the class-conditional TimeVQVAE and other baselines, demonstrating better generation quality and strong representation disentanglement without using class labels.
In conclusion, the proposed TimeDP model effectively tackles multi-domain time series generation by using domain prompts and prototypes. It outperforms existing methods, offering better in-domain quality and strong performance on unseen domains. This approach sets a new benchmark for time series generation and can serve as a basis for future research, particularly in prototype-based learning and domain adaptation. Future work could improve scalability and explore its use in more complex applications.
Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. Don’t Forget to join our 65k+ ML SubReddit.
Recommend Open-Source Platform: Parlant is a framework that transforms how AI agents make decisions in customer-facing scenarios. (Promoted)
The post TimeDP: A Multi-Domain Time Series Diffusion Model with Domain Prompts appeared first on MarkTechPost.
