Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models

This is the official repository for the paper Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models accepted by TMLR . In combination with (conditional) diffusion and state-space models, we put forward diverse algorithms, particualary, we propose the generative model $SSSD^{S4}$, which is suited to capture long-term dependencies and demonstrates state-of-the-art results in time series across diverse missing scenarios and datasets.

Datasets and experiments

Visit the source directory to get datasets download and experiments reproducibility instructions. (here is an example of the feature sampling approach for the datasets with large number of channels )

Our proposed $SSSD^{S4}$ model architecture:

$SSSD^{S4}$ robustness on diverse scenarios:

Random Missing

Missing not at random

Black-out missing

Forecast

Please cite our publication if you found our research to be helpful.

@article{
lopez alcaraz2022diffusionbased,
title={Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models},
author={Juan Lopez Alcaraz and Nils Strodthoff},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2022},
url={https://openreview.net/forum?id=hHiIbk7ApW},
}

Acknowledgments

We would like thank the authors of the the S4 model for releasing and maintaining the source code for Structured State Space Models. Similarly, our proposed model code builds on the implementation provided by DiffWave.

jdariasl/SSSD