On the Constrained Time-Series Generation Problem
Published in Advances in Neural Information Processing Systems (NeurIPS 2023), 2023
Recommended citation: Coletta, A., Gopalakrishnan, S., Borrajo, D., & Vyetrenko, S. (2023). On the Constrained Time-Series Generation Problem. In Advances in Neural Information Processing Systems (NeurIPS). https://proceedings.neurips.cc/paper_files/paper/2023/file/a8b5ebae9b3eb16b0811e5cd8d20cf4c-Paper-Conference.pdf
Abstract
Generating synthetic time series satisfying strict physical, financial, and domain constraints is critical for real-world simulation and privacy-preserving data sharing. We formalize the constrained time-series generation problem and propose a novel architecture marrying score-based diffusion models with projected gradient methods to strictly satisfy non-convex constraint sets.
Recommended citation: Coletta, A., Gopalakrishnan, S., Borrajo, D., & Vyetrenko, S. (2023). On the Constrained Time-Series Generation Problem. In Advances in Neural Information Processing Systems (NeurIPS).