The figure depicts a framework of the training stages incorporating wavelet augmentations, which involve the concatenation of
the look-back window and the forecasting horizon prior to transformation and augmentation. Batch sampling of the generated synthetic data is conducted according to a predefined hyperparameter called the sampling rate. These batches are subsequently used to split the data into the look-back window and target horizon, after which they are concatenated with the original data. Wavelet augmentations are Wavelet Masking (WaveMask) and Wavelet Mixing (WaveMix). These techniques utilize the discrete wavelet transform (DWT) to obtain wavelet coefficients (both approximation and detail coefficients) by breaking down the signal and adjusting these coefficients, in line with modifying frequency components across different time scales.
WaveMask selectively eliminates specific wavelet coefficients at each decomposition level, thereby introducing variability in the augmented data. Conversely, WaveMix exchanges wavelet coefficients from two distinct instances of the dataset, thereby enhancing the diversity of the augmented data.
To the best of our knowledge, this is the first study to conduct extensive experiments on multivariate time series using Discrete Wavelet Transform as an augmentation technique.
Overview
Tables present comparisons between our methods and baselines in terms of the metrics Mean Squared Error (MSE) and Mean Absolute Error (MAE). The best result is indicated in bold, while the second most favorable outcome is underlined.
You can change percentage to down-sample training dataset for ablation study.
Citation
If you find the code useful, please cite our paper:
@misc{waveaug2024,
title={Wave-Mask/Mix: Exploring Wavelet-Based Augmentations for Time Series Forecasting},
author={Dona Arabi and Jafar Bakhshaliyev and Ayse Coskuner and Kiran Madhusudhanan and Kami Serdar Uckardes},
year={2024},
eprint={2408.10951},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2408.10951},
}
Please remember to cite all the datasets and compared methods if you use them in your experiments.
Acknowledgements
We would like to express our gratitude to Zheng et al. for providing datasets used in this project. Additionally, we also acknowledge Chen et al. and Zhang et al. for their code frameworks, which served as the foundation for our codebase.
