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

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Human activity recognition (HAR) using wearable sensors faces a performance-labeling trade-off.
• Fully supervised methods require extensive, costly labeled data.
• Unsupervised methods reduce labeling but often yield lower performance.

Purpose of the Study:

• Investigate the full spectrum of supervision for wearable-based HAR.
• Focus on novel approaches minimizing labeling while maintaining accuracy.
• Compare traditional supervised, unsupervised, weakly supervised, multi-task, self-supervised, and weakly self-supervised learning.

Main Methods:

• Developed and compared six learning paradigms: fully supervised, unsupervised, weakly supervised, multi-task, self-supervised, and a novel weakly self-supervised framework.
• Leveraged domain knowledge and minimal labeled data in the proposed framework.
• Conducted experiments on benchmark datasets for wearable-based HAR.

Main Results:

• Weakly supervised methods achieved performance comparable to fully supervised approaches with reduced supervision.
• The multi-task framework improved performance via knowledge sharing.
• The novel weakly self-supervised approach demonstrated high efficiency using only 10% labeled data.

Conclusions:

• Different learning paradigms offer complementary strengths for HAR based on data availability.
• The novel weakly self-supervised framework is a promising solution for practical HAR with limited labeled data.