Core Concepts
Data augmentation techniques have varying impacts on time-series classification performance, highlighting the importance of judicious selection.
Abstract
The study delves into the significance of Data Augmentation (DA) in Time Series Classification (TSC), emphasizing its role in enhancing model robustness and diversifying datasets. The research conducts an extensive empirical study and survey to dissect DA methodologies within TSC, identifying over 60 unique techniques categorized into five principal echelons. An evaluation across 8 UCR time-series datasets using ResNet reveals varying efficacies of DA strategies, with some significantly improving model performance while others compromising it. Dataset attributes are found to impact the success of DA techniques, leading to precise recommendations for practitioners.
Structure:
Introduction to Data Augmentation in Time-Series Classification
Methodology: Extensive Literature Review and Empirical Assessment
Taxonomy of DA Techniques: Transformation-Based, Pattern-Based, Generative, Decomposition-Based, Automated DA
Evaluation Across UCR Datasets Using ResNet: Accuracy, Method Ranking, Residual Analysis
Discussion on Impact and Effectiveness of Different DA Techniques
Stats
Random Permutation boasts an average accuracy of 89.32 ± 11.65%
Permutation achieves an accuracy pinnacle of 89.20 ± 11.95%
Rotation inadvertently compromises model performance with an accuracy of 84.75 ± 12.92%
Quotes
"The benefits of augmentation are not universally applicable."
"Dataset attributes exert a profound impact on the success of DA techniques."