Time Series Data Augmentation for Neural Networks by Time Warping with a Discriminative Teacher (2004.08780v1)

Abstract: Neural networks have become a powerful tool in pattern recognition and part of their success is due to generalization from using large datasets. However, unlike other domains, time series classification datasets are often small. In order to address this problem, we propose a novel time series data augmentation called guided warping. While many data augmentation methods are based on random transformations, guided warping exploits the element alignment properties of Dynamic Time Warping (DTW) and shapeDTW, a high-level DTW method based on shape descriptors, to deterministically warp sample patterns. In this way, the time series are mixed by warping the features of a sample pattern to match the time steps of a reference pattern. Furthermore, we introduce a discriminative teacher in order to serve as a directed reference for the guided warping. We evaluate the method on all 85 datasets in the 2015 UCR Time Series Archive with a deep convolutional neural network (CNN) and a recurrent neural network (RNN). The code with an easy to use implementation can be found at https://github.com/uchidalab/time_series_augmentation .

• The paper's main contribution is a deterministic guided warping method that leverages DTW and a discriminative teacher for effective time series data augmentation.
• The methodology aligns features to preserve temporal relationships, outperforming random transformation techniques on the UCR Time Series Archive datasets.
• Results show significant accuracy improvements in CNNs and RNNs, demonstrating its potential in sensor analytics, bioinformatics, and other time-dependent fields.

Overview of Time Series Data Augmentation Using Guided Warping

This paper presents a method to address the challenges of limited data availability in time series classification by introducing a novel data augmentation approach termed "guided warping." This method leverages Dynamic Time Warping (DTW) and its extension, shapeDTW, to enhance the training of neural networks by expanding small datasets with synthesized, yet realistic, data samples.

Methodology

The authors propose a deterministic approach to data augmentation, contrasting with traditional random transformation methods. Guided warping utilizes DTW to align and warp time series samples by aligning the features of a sample pattern to the temporal structure of a reference pattern. This alignment preserves temporal relationships, facilitating well-distributed and meaningful augmentation without baseless randomness, ensuring the augmented data enriches the original dataset's feature space.

Moreover, the paper introduces the concept of a "discriminative teacher" for optimizing reference pattern selection. This involves a nearest centroid classifier to identify reference patterns that maximize the discriminative distance between classes, enhancing the efficacy of the guided warping process.

Results and Comparisons

The proposed method was evaluated on the 85 datasets from the UCR Time Series Archive using both CNNs and RNNs. It demonstrated significant accuracy improvements, particularly with convolutional architectures, where VGG-based models showed a marked performance increase. For instance, data types such as simulated and motion capture saw impressive accuracy gains.

The guided warping approach mitigated the creation of unrealistic samples often generated by random methods, maintaining the integrity of the time series data's original feature distribution. With the discriminative teacher strategy, the method further outperformed simpler pattern mixing techniques like SPAWNER and wDBA ASD, which either generated overly similar patterns or poorly distributed samples.

Implications and Future Directions

Guided warping presents a promising augmentation method that aligns well with modern neural network architectures. Its effectiveness in augmenting small datasets without diminishing data quality is significant for fields like sensor data analytics and bioinformatics, where dataset sizes are inherently limited.

The research opens pathways for exploring similar augmentation techniques in multivariate time series data and integration with more advanced neural architectures, such as transformers in time series analysis. Future work could also explore the dynamic adaptation of the discriminative teacher selection process to further refine the augmentation strategy during training.

Given its contribution to improving classifier performance, this approach could see broader applications across various domains where time-dependent features are critical. The continued development and adaptation of such augmentation methods will likely be crucial as data-driven models increasingly find themselves applied to diverse and complex real-world datasets.