Multi-Resolution Diffusion Models for Time Series Forecasting

TimeGrad [ICML'21] CSDI [NeurIPS'21] TimeDiff [ICML'23] generate time series directly from random vectors Motivation how to better utilize structural properties in time series? multi-resolution temporal structure in time series • seasonal-trend decomposition: extract the seasonal and trend components • the coarser temporal patterns can be used to help modeling the finer patterns how to use multi-resolution analysis in time series diffusion models? • coarser trends are generated first and the finer details are then progressively added. • decompose the denoising objective into several sub-objectives, each corresponds to a particular resolution.