Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches

This paper explores methods to reduce the reliance on labeled data in human activity recognition (HAR) using wearable sensors. It investigates various machine learning paradigms, including supervised, unsupervised, weakly supervised, multi-task, and self-supervised learning. The core contribution is a novel weakly self-supervised learning framework that combines domain knowledge with minimal labeled data. The experimental results demonstrate that the proposed weakly supervised methods can achieve performance comparable to fully supervised approaches while significantly reducing supervision requirements. The multi-task framework also shows performance improvements through knowledge sharing. This research is significant because it addresses the practical challenge of limited labeled data in HAR, making it more accessible and scalable.

• •Weakly supervised learning can achieve comparable performance to fully supervised learning with less labeled data.
• •Multi-task learning can improve performance through knowledge sharing between related tasks.
• •Self-supervised learning, especially when combined with domain knowledge, offers a promising avenue for reducing label dependency.