Label-preserving data augmentation for mobile sensor data

Data augmentation is important for training neural networks, especially when there is not enough data to train a network well. However, data augmentation that results in the loss of label information may reduce the performance of the model. Most conventional data augmentation methods have been developed for image- or sound-related tasks, in which case the label information of the augmented data is easily and intuitively verified by human observation. However, in the case of sensor signals, it is difficult to recognize whether there is a change in the label information of the augmented data. We propose a systematic data augmentation method to maximize the performance by automatically finding the range of augmentation that preserves the labels of the augmented data. The experimental results show that the proposed method to extract the label-preserving range is practical and that the retrained model using data augmented within this range improves the performance by at least 5% without the need to further optimize the model architecture.

Automated summary

Data augmentation is essential for training neural networks when data is limited, but it may lead to performance reduction if label information is lost. Traditional augmentation methods for image or sound tasks are easily validated by observation, but for sensor signals, it is difficult to preserve label information. Our proposed systematic augmentation method automatically finds the range that maintains label information, resulting in a performance improvement of at least 5% without further architectural optimization.