Source-linked AI summary
Data augmentation using synthetic data for time series classification with deep residual networks
Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, Pierre-Alain Muller
TL;DR
Small TSC datasets can cause poor CNN generalization, while suitable augmentation methods remain underexplored. The paper applies DTW-based weighted DBA augmentation to a deep ResNet and evaluates it on the UCR benchmark. Augmentation sharply improves some datasets, and an ensemble reduces its negative effects while preserving gains.
Problem
Small TSC datasets can limit CNN generalization, while time-series data augmentation has been less thoroughly investigated than image augmentation.
Method
The paper applies a DTW-based weighted DBA synthetic-series generation method to a validated deep ResNet for TSC.
Results
Augmentation drastically improves accuracy on some datasets, while combining augmented and unaugmented ResNets reduces rare negative effects and preserves gains.
Takeaways & Limitations
The experiments indicate that synthetic time series can help ResNets learn useful time-invariant classification features when training data are limited.
Takeaways & Limitations
The study does not test how imbalanced training classes affect generalization and leaves the generation-count hyperparameter for future investigation.
Abstract
from arXiv · showhide
Data augmentation in deep neural networks is the process of generating artificial data in order to reduce the variance of the classifier with the goal to reduce the number of errors. This idea has been shown to improve deep neural network's generalization capabilities in many computer vision tasks such as image recognition and object localization. Apart from these applications, deep Convolutional Neural Networks (CNNs) have also recently gained popularity in the Time Series Classification (TSC) community. However, unlike in image recognition problems, data augmentation techniques have not yet been investigated thoroughly for the TSC task. This is surprising as the accuracy of deep learning models for TSC could potentially be improved, especially for small datasets that exhibit overfitting, when a data augmentation method is adopted. In this paper, we fill this gap by investigating the application of a recently proposed data augmentation technique based on the Dynamic Time Warping distance, for a deep learning model for TSC. To evaluate the potential of augmenting the training set, we performed extensive experiments using the UCR TSC benchmark. Our preliminary experiments reveal that data augmentation can drastically increase deep CNN's accuracy on some datasets and significantly improve the deep model's accuracy when the method is used in an ensemble approach.
1 Introduction
Small TSC datasets can limit CNN generalization, motivating time-series-specific data augmentation. The paper proposes a DTW-based augmentation approach for ResNet and reports large gains on some datasets, with an ensemble reducing adverse effects.
- Twenty UCR datasets have 50 or fewer training instances, far smaller than typical successful computer-vision training sets.
- CNNs can generalize poorly on small datasets such as DiatomSizeReduction despite strong performance from 1-NN with DTW.
- Time-series augmentation remains limited because transformations such as image-style rotations cannot be easily verified as label-preserving.
- The paper applies a DTW-based augmentation technique to improve a deep ResNet for time-series classification.
- Augmentation can sharply improve CNN accuracy on some datasets but slightly reduce it on others; an ensemble reduces the rare negative effect.
2 Related work
Prior TSC augmentation work centers on slicing windows and warping, adapting ideas from image augmentation but introducing classification and information-preservation concerns for time series.
- Slicing windows are the most-used TSC augmentation method and were originally adapted from image cropping for deep CNNs.
- For time series, cropping may remove discriminative information, so the whole-series label is not guaranteed to remain valid.
- Other proposed generic transformations include jittering, scaling, warping, and permutation, alongside wearable-sensor-specific augmentation.
- Warping-window methods generate synthetic series through temporal warping but require classifying subsequences and combining them by majority voting.
3 Method
The method combines a validated deep ResNet with DTW-based synthetic-series generation using weighted DBA. The study also notes untested class-imbalance effects and a generation-count hyperparameter.
- Architecture: The study uses an already validated ResNet so performance changes can be attributed to data augmentation, while targeting a deep architecture suited to large TSC datasets.
- Architecture: The network maps a variable-length univariate series to class probabilities through three residual blocks, global average pooling, and softmax classification.
- Architecture: Each residual block uses three 1-D convolutions with batch normalization, ReLU activations, residual additions, and increasing filter counts.
- Data augmentation: Synthetic series are generated with weighted DTW Barycentric Averaging, which can create infinitely many averages by varying weights.
- Data augmentation: The weighting procedure initializes a random training series at 0.5, finds five DTW nearest neighbors, and randomly selects two neighbors.
- Data augmentation: Weighted DBA was selected for prior use in augmentation, although alternatives such as soft-DTW could replace it.
- Data augmentation: The study does not test class imbalance and sets generated-series count to twice the size of the most represented class.
4 Results
Experiments on the UCR archive show that DTW-based augmentation can substantially improve ResNet accuracy on some datasets, while an ensemble reduces harmful cases. The evaluation used consistent initialization across approaches.
- Experimental setup: The method was evaluated on the UCR archive, a large benchmark containing datasets with varied characteristics.The experiments used more than 60 GPUs, and compared approaches received the same parameter initializations.
- Effect of data augmentation: Accuracy increased from 30% to 96% on DiatomSizeReduction when data augmentation was applied.This dataset had only 16 training instances, and its improvement was attributed to generating synthetic time series.
- Effect of data augmentation: Data augmentation drastically improved accuracy overall while causing only a small negative impact on some datasets.The results indicate that augmentation did not lead to a significant decrease in accuracy in the reported comparison.
- Ensemble results: The ensemble reduced datasets harmed by augmentation from 30 to 21, with a significant Wilcoxon signed rank result (p-value < 0.0005).It averaged posterior probabilities from ResNets trained with and without augmentation.
5 Conclusion
The paper concludes that DTW-based synthetic time series can mitigate overfitting in small TSC datasets and help ResNets learn useful time-invariant features. Future work will test weighting schemes and predict when augmentation is beneficial.
- Conclusion: DTW-based augmentation with a weighted DBA algorithm mitigated overfitting in small time series datasets.The conclusion links the approach to providing enough time series for ResNets to learn useful classification features.
- Future work: Future work will test alternative DTW weighting schemes and predict when augmentation benefits particular datasets.This identifies the method’s current scope as not yet including dataset-specific benefit prediction.