Data Augmentation for Time-Series Classification: a Comprehensive Survey

Data Augmentation (DA) for Time Series Classification (TSC) is a common technique in machine learning to increase the number of training samples, which enhances model performance, enriches the dataset variety, and helps mitigate overfitting. Nonetheless, this technique is currently faced with challenges characterized by incomplete reviews, ambiguous taxonomies, insufficient evaluations, and user-unfriendly tools. This study undertakes a detailed exploration of DA for TSC. We first conducted a thorough review of the developments in the field of DA for TSC over the past 10 years since existing surveys on DA for TSC are not comprehensive enough. Our efforts encompassed gathering more than 60 distinct DA techniques from a pool over 100 research papers. This endeavor culminated in the creation of an innovative taxonomy exclusively tailored to DA within the TSC domain. The taxonomy organizes methods into five main categories: Transformation-Based, Pattern-Based, Generative, Decomposition-Based, and Automated Data Augmentation. This classification serves as a sturdy reference for researchers when choosing methods. In addition, since there is a lack of comprehensive and detailed evaluations of popular data augmentation methods, we conduct a comprehensive assessment. More than 15 DA methods were tested on 8 UCR time-series datasets using the ResNet and deploying a multi-metric evaluation strategy that includes Accuracy, Method Ranking, and Residual Analysis, the outcome was a baseline accuracy of 88.94 +- 11.83%. Findings highlighted the variable effectiveness of DA methods, for instance, methods like Permutation enhanced performance while Rotation decreased accuracy. Dataset properties also profoundly influence DA efficacy, we give users accurate and practical advice based on our experimental results to guide them in choosing the most appropriate DA methods for different data characteristics.