Towards a Unified Generative Model for Scarce Time Series with Domain Experts

In three linesTimeMoDE, a framework combining Diffusion Transformers and Mixture-of-Experts, generates realistic time series under data scarcity. Pre-trained on multi-domain datasets, it uses Domain Prompts to condition expert assignment and incorporates diffusion timestep signals for adaptive denoising. Outperforms existing methods in few-shot generation settings.

Summary generated by Claude — human-verified