[Paper Review] Adaptive Weighting Scheme for Automatic Time-Series Data Augmentation

Elizabeth Fons, P Dawson+3 more|arXiv (Cornell University)|Feb 16, 2021

Time Series Analysis and Forecasting23 references17 citations

TL;DR

This paper proposes two sample-adaptive data augmentation methods for time-series classification: W-Augment, which learns to weight the loss contribution of each augmented sample, and α-trimmed Augment, which selects the most effective transformations based on predicted loss. The methods achieve over 50% higher annualized returns on a financial dataset and outperform state-of-the-art models on over half of UCR time-series datasets, demonstrating superior generalization through adaptive, learnable augmentation policies.

ABSTRACT

Data augmentation methods have been shown to be a fundamental technique to improve generalization in tasks such as image, text and audio classification. Recently, automated augmentation methods have led to further improvements on image classification and object detection leading to state-of-the-art performances. Nevertheless, little work has been done on time-series data, an area that could greatly benefit from automated data augmentation given the usually limited size of the datasets. We present two sample-adaptive automatic weighting schemes for data augmentation: the first learns to weight the contribution of the augmented samples to the loss, and the second method selects a subset of transformations based on the ranking of the predicted training loss. We validate our proposed methods on a large, noisy financial dataset and on time-series datasets from the UCR archive. On the financial dataset, we show that the methods in combination with a trading strategy lead to improvements in annualized returns of over 50$\%$, and on the time-series data we outperform state-of-the-art models on over half of the datasets, and achieve similar performance in accuracy on the others.

Motivation & Objective

• To address the lack of automated, adaptive data augmentation for time-series data, especially given small and noisy datasets.
• To develop sample-adaptive policies that dynamically weight or select augmentations based on individual sample performance.
• To improve model generalization and robustness in time-series classification by learning optimal transformation strategies.
• To validate the proposed methods on real-world financial data and benchmark UCR time-series datasets.
• To demonstrate that adaptive augmentation outperforms fixed policies like RandAugment in both accuracy and financial performance.

Proposed method

• W-Augment introduces a trainable weight vector that assigns dynamic, sample-specific weights to each augmented sample's loss contribution during backpropagation.
• α-trimmed Augment selects a subset of transformations per sample based on their predicted training loss, discarding the worst-performing ones.
• A single distortion magnitude parameter M is used to control the strength of all augmentations uniformly, reducing hyperparameter search space.
• The methods are integrated with InceptionTime, an ensemble of five 1D CNNs, and trained end-to-end with joint optimization of network and augmentation weights.
• The approach extends RandAugment to time-series by applying it with equal probability to mini-batches, but adds sample-adaptive learning to improve performance.
• Hyperparameter tuning selects M from {1,5,10,15,20} using validation accuracy, with the final model averaging results across five data splits.

Experimental results

Research questions

• RQ1 Can sample-adaptive data augmentation improve time-series classification performance compared to fixed, non-adaptive policies?
• RQ2 Does learning to weight or select augmentations based on individual sample loss lead to better generalization on time-series data?
• RQ3 Can adaptive augmentation significantly improve financial time-series prediction, especially in terms of return on investment?
• RQ4 How does the number of included transformations affect performance, and does adaptive selection mitigate sub-optimal augmentations?
• RQ5 Can the proposed methods outperform RandAugment and other state-of-the-art models on standard UCR time-series benchmarks?

Key findings

• On the S&P500 financial dataset, the proposed augmentation methods combined with a trading strategy increased annualized returns by over 50%.
• On UCR time-series datasets, W-Augment outperformed the baseline on 6 out of 12 datasets and matched or exceeded it on the remaining 6.
• With 8 transformations, W-Augment won or drew 8 out of 12 datasets, while RandAugment won only 5, indicating superior handling of sub-optimal augmentations.
• The α-trimmed Augment method achieved better performance than the baseline on 7 out of 12 datasets and matched it on 5, showing effectiveness in selective augmentation.
• Learned weights in W-Augment showed consistent behavior across InceptionTime’s five models, with higher weights assigned to original samples and dynamic adaptation across datasets.
• Validation accuracy improved with more transformations when using W-Augment and α-trimmed Augment, suggesting scalability and robustness to augmentation set size.