Time-Series AutoAugment: Data Augmentation Policy Search for Long-Term Forecasting

Liran Nochumsohn, Omri Azencot

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Time-series forecasting, data augmentation

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Data augmentation policy search framework designed for time-series forecasting.

Abstract:

Data augmentation is a popular regularization for addressing overfitting issues
of neural networks. Recently, automatic augmentation showed strong results on
image classification tasks. However, less attention had been given to automatic
augmentation of time-series problems such as long-term forecasting. Toward
bridging this gap, we propose an efficient, effective, and easy-to-code time-series
automatic augmentation method we refer to as TSAA. We solve the associated
bilevel optimization problem in two steps: a partial train of the non-augmented
model for a few epochs and an iterative split process. The iterative process alternates
between finding a good augmentation policy via Bayesian optimization and fine-
tuning the model while pruning poor runs. Our method is evaluated extensively
on challenging univariate and multivariate forecasting benchmark problems. Our
results indicate that TSAA outperforms several strong baselines in most cases,
suggesting it should be incorporated into prediction pipelines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1405