Abstract
The advent of wearable technologies ushered in an era of abundant time series data, offering profound insights into human health and behavior. However, the full utilization of such data was hindered by challenges such as the scarcity of labeled datasets and the need for privacy in sensitive domains like personal health tracking. To address these challenges, this paper introduces Hydra-TS, a multi-agent generative adversarial network. Hydra-TS uniquely excels in optimizing multiple objectives concurrently. Hydra-TS offers a spectral representation for time series data. Here, a single generator is pitted against a variable number of discriminators to create multivariate synthetic data that are realistic, useful for classification, and privacy preserving. Using a one-month dataset of real-world, in-the-wild smartwatch data containing 5,271,143 labeled activity instances for 10 participants, we demonstrated that Hydra-TS yielded a superior Area under the Radar Chart value (AuRC=.72) in comparison with the original data and three baselines methods. We also verified that activity recognition performance was improved using Hydra-TS as a vehicle for data augmentation, improving the f1 score by as much as 130.54%. Hydra-TS's effectiveness underlines its potential to facilitate research and applications in areas where data scarcity and privacy issues are prevalent.