Data Augmentation Strategies for Human Activity Data Using Generative Adversarial Neural Networks

Previous studies have shown that available benchmark datasets from the field of Human Activity Recognition are of limited use for Deep Learning applications. This can be traced back to issues in the quality, the scope, as well as in the variability of the datasets. These limitations often lead to overfitting of networks and thus to results that are only conditionally generalizable. One way to counteract this problem is to extend the data by using data augmentation techniques. This paper presents an algorithm and compares two augmentation strategies: (1) user-wise augmentation and (2) fold-wise augmentation to extend the size of a dataset here shown on the PAMAP2 dataset, with an arbitrary number of synthetic samples. These synthesized data resemble the user- and activity-specific characteristics and fit seamlessly into the dataset. They are created by a recurrent Generative Adversarial Network, with both the generator and discriminator modeled by a set of LSTM cells to produce the synthetic time-series data. In our evaluation, we trained four DeepConvLSTM models with supervised learning, three times with a LOSO cross-validation: one baseline model and two times with additional data but different augmentation strategies, as well as one model without cross-validation that monitors the synthesized data quality. The compared augmentation strategies demonstrate the impact as well as the generalized nature of the augmented data. By increasing the size of the dataset by factor 5, we improved the F1-Score by 11.0% with strategy (1) and 5.1% with strategy (2).