Latent Laplace Diffusion for Irregular Multivariate Time Series

Researchers introduced LLapDiff, a diffusion-based method for irregular multivariate time series.

It models targets as latent trajectories and uses Laplace-domain parameterization with complex poles to handle uneven timestamps.

The approach avoids costly sequential solvers while preserving temporal structure in the data.

Results show stronger long-horizon forecasting performance and effective missing-value imputation.