TimeMoDE Generates Realistic Time Series from Scarce Data

Generating realistic synthetic time series data is crucial for many real-world applications, but most existing generative models require extensive training data. This reliance on abundant data severely limits their utility in scenarios where data is scarce, which is a common problem across various domains. To address this, researchers have introduced TimeMoDE, a novel framework designed for time series generation under data scarcity. TimeMoDE integrates Diffusion Transformers with a Mixture-of-Experts architecture. This combination allows it to exploit both domain adaptability and an awareness of the diffusion stage, making it highly effective even with limited data. TimeMoDE is pre-trained on a large collection of multi-domain datasets to learn both general temporal representations and domain-specific information, which aids generalization during fine-tuning. It uses "Domain Prompts" to guide expert assignment for noisy tokens and incorporates diffusion timestep signals to equip experts with awareness of time series degradation, facilitating adaptive denoising. Extensive experiments confirm that TimeMoDE outperforms current methods in diverse low-data settings, establishing a new paradigm for few-shot time series generation.