Hydra-TS: Enhancing Human Activity Recognition with Multi-Objective Synthetic Time Series Data Generation.

Area of Science:

• Wearable technology
• Machine learning
• Data science

Background:

• Wearable devices generate vast amounts of time series data for health and behavior insights.
• Challenges include limited labeled datasets and privacy concerns in sensitive health tracking.
• Existing methods struggle to optimize multiple objectives like realism, utility, and privacy simultaneously.

Purpose of the Study:

• Introduce Hydra-TS, a multi-agent generative adversarial network (MAGAN).
• Address limitations in synthetic time series data generation for wearable technology.
• Optimize for data realism, classification utility, and privacy preservation concurrently.

Main Methods:

• Developed Hydra-TS, a MAGAN with one generator and multiple discriminators.
• Utilized a spectral representation for time series data.
• Trained on a one-month smartwatch dataset with over 5 million labeled activity instances from 10 participants.

Main Results:

• Hydra-TS achieved a superior Area under the Radar Chart (AuRC) value of 0.72 compared to original data and baselines.
• Data augmentation with Hydra-TS improved activity recognition F1 scores by up to 130.54%.
• Demonstrated effectiveness in generating realistic, useful, and privacy-preserving synthetic multivariate time series data.

Conclusions:

• Hydra-TS successfully generates high-quality synthetic time series data.
• The method significantly enhances activity recognition performance through data augmentation.
• Hydra-TS offers a promising solution for research and applications facing data scarcity and privacy issues.