Improving smartphone based transport mode recognition using generative adversarial networks

Wearable devices such as smartphones and smartwatches are widely used and record a significant amount of data. Labelling this data for human activity recognition is a time-consuming task, therefore methods which reduce the amount of labelled data required to train accurate classifiers are important. Generative Adversarial Networks (GANs) can be used to model the implicit distribution of a dataset. Traditional GANs, which only consist of a generator and a discriminator, result in networks able to generate synthetic data and distinguish real from fake samples. This adversarial game can be extended to include a classifier, which allows the training of the classification network to be enhanced with synthetic and unlabelled data. The network architecture presented in this paper is inspired by SenseGAN [1], but instead of generating and classifying sensor-recorded time-series data, our approach operates with extracted features, which drastically reduces the amount of stored and processed data and enables deployment on less powerful and potentially wearable devices. We show that this technique can be used to improve the classification performance of a classifier trained to recognise locomotion modes based on recorded acceleration data and that it reduces the amount of labelled training data necessary to achieve a similar performance compared to a baseline classifier. Specifically, our approach reached the same accuracy as the baseline classifier up to 50% faster and was able to achieve a 10% higher accuracy in the same number of epochs.