TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series

Alexandre Drouin, Étienne Marcotte, Nicolas Chapados (2023). TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series. (Preprint)

We introduce a new model for multivariate probabilistic time series prediction, designed to flexibly address a range of tasks including forecasting, interpolation, and their combinations. Building on copula theory, we propose a simplified objective for the recently-introduced transformer-based attentional copulas (TACTiS), wherein the number of distributional parameters now scales linearly with the number of variables instead of factorially. The new objective requires the introduction of a training curriculum, which goes hand-in-hand with necessary changes to the original architecture. We show that the resulting model has significantly better training dynamics and achieves state-of-the-art performance across diverse real-world forecasting tasks, while maintaining the flexibility of prior work, such as seamless handling of unaligned and unevenly-sampled time series.

Installation

You can install TACTiS with pip:

pip install tactis

To install the research version of TACTiS to replicate/run experiments as in the paper:

pip install tactis[research]

Note: tactis has been currently tested with Python 3.10.8.

Details

train.py can be used to train the TACTiS-2 model for a dataset. The arguments in train.py can be used to specify the dataset, the training task (forecasting or interpolation), the hyperparameters of the model and a whole range of other training options. More features such as interfaces to evaluate the model and obtain metrics, as well as demo notebooks will be added to this codebase soon.

For an implementation of the original version of TACTiS, please see here.

Citing this work

Please use the following Bibtex entry to cite TACTiS-2.

@misc{ashok2023tactis2,
      title={TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series}, 
      author={Arjun Ashok and Étienne Marcotte and Valentina Zantedeschi and Nicolas Chapados and Alexandre Drouin},
      year={2023},
      eprint={2310.01327},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}