Fig 1.
Examples of random transformation-based data augmentation on the OliveOil dataset.
The dotted blue lines are the original patterns and the solid red lines are the generated patterns.
Fig 2.
Taxonomy of time series data augmentation.
Fig 3.
The architectures of the neural networks used to evaluate the data augmentation methods.
Table 1.
Comparative results for magnitude domain transformation-based data augmentation methods.
Table 2.
Comparative results for time domain transformation-based data augmentation methods.
Table 3.
Comparative results for pattern mixing-based data augmentation methods.
Fig 4.
Visualization of the GunPoint dataset using PCA and the compared augmentation methods.
The solid shapes are the original time series, the hollow shapes are the generated time series, and the colors indicate different classes.
Fig 5.
Spearman’s Rank Correlation Coefficients between change in accuracy (Δ Acc) and various dataset characteristics.
The top row in red is the change in accuracy and the subsequent rows are the correlations.
Table 4.
Algorithm comparison with average augmentation time per dataset and tunable parameters.
Table 5.
Data augmentation recommendations for data type and model type.