Domain Adaptation in Children Activity Recognition.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Training predictive models for wireless health faces challenges in adapting to new individuals or groups.
• Data distribution differences between real-world deployment and training data hinder model generalization.

Purpose of the Study:

• To address the challenge of model adaptation in wireless health by employing deep domain adaptation techniques.
• To transfer a model trained on one data distribution (adults) to a different distribution (children) for activity recognition.

Main Methods:

• Utilized deep domain adaptation to transfer a pre-trained activity recognition model.
• Evaluated the model's performance on a new dataset (children) without requiring new labels.
• Compared the performance against a direct supervised baseline.

Main Results:

• Directly applying an adult-trained model to children resulted in a 25.2% loss in F1-score compared to a supervised baseline.
• The proposed domain adaptation approach reduced the performance gap to only a 9% loss in F1-score.

Conclusions:

• Deep domain adaptation is effective in overcoming data distribution shifts for wireless health models.
• This approach significantly improves the performance of activity recognition models when applied to new populations, like children.