Efficient time series data augmentation on medical wearable sensors

What is it about?

Recent advancements in wearable sensor technology have led to a surge in Time Series (TS) data capturing behavioral and physiological information for medical applications. However, the complexity and time demands of data collection limit dataset sizes, affecting the performance of machine learning models. Data Augmentation (DA) offers a promising solution by generating synthetic data to overcome these challenges. While DA is well established in image domains, its application to TS classification remains less mature, facing issues such as a focus on univariate methods, a lack of guidance in selecting techniques, and uncertainty about the optimal quantity of synthetic data. In this paper, we present a comprehensive survey and experimental evaluation of Time Series Data Augmentation (TSDA) techniques, proposing an updated taxonomy across three categories: Random Transformation (RT), Pattern Mixing (PM), and Generative Models (GM). We evaluate 12 TSDA methods on diverse medical-related datasets, including OPPORTUNITY, HAR, DEAP, BVDB, and PMDB. Our results highlight that simple RT methods are consistently the most effective in improving classification performance compared to no augmentation.

Featured Image

Why is it important?

The paper will give a complete insight into modern data augmentation techniques and their applicability to spatio-temporal datasets (specifically human activities data). The paper also provides a thorough comparative analysis of the DA techniques on the Time Series dataset.

Perspectives