Reproduce “Empirical Study of Mix-based Data Augmentation Methods in Physiological Time Series Data” on OpenResearch

About this paper

Peikun Guo, Huiyuan Yang, Akane Sano

Researchers from Rice University conducted the first comprehensive empirical study of mix-based data augmentation methods for physiological time series data, demonstrating their consistent performance improvement and robustness across diverse datasets. The study found these methods lead to more discriminative feature representations and effectively mitigate class imbalance, outperforming traditional augmentation techniques.

Key insights

• Mix-based augmentations provide robust performance and consistent accuracy gains with minimal reliance on expert knowledge or extensive parameter tuning, addressing the data-dependency of traditional methods.
• Unlike traditional methods, mix-based augmentations generally do not compromise the integrity of delicate physiological patterns, yielding stable and reliable improvements.
• Mix-based methods enhance feature representation, leading to more discriminative class embeddings, and are effective in mitigating class imbalance when combined with class-balanced sampling.

Summary by alphaXiv.