Towards User-Generalizable Wearable-Sensor-Based Human Activity Recognition: A Multi-Task Contrastive Learning...

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Human Activity Recognition (HAR) using wearable sensors is crucial for personalized health and ubiquitous computing.
• Current deep learning HAR models struggle with user-level generalization, hindering real-world application.
• A need exists for HAR models that perform reliably across diverse, unseen users.

Purpose of the Study:

• To develop a novel multi-task contrastive learning framework for enhanced user-level generalization in HAR.
• To improve the robustness and scalability of HAR systems for real-world deployment.
• To investigate the impact of joint activity classification and user-aware contrastive learning on HAR performance.

Main Methods:

• Proposed a multi-task contrastive learning framework combining activity classification and supervised contrastive objectives.
• Utilized both activity and user labels to create semantically rich contrastive pairs.
• Employed a user-agnostic inference strategy for testing on unseen users.

Main Results:

• Achieved comparable results to supervised and self-supervised baselines on three public HAR datasets using cross-user evaluation.
• Demonstrated the effectiveness of multi-task training and user-aware contrastive supervision through ablation studies.
• Showcased improved representation learning for better generalization across users.

Conclusions:

• The proposed multi-task contrastive learning framework effectively enhances user-level generalization in HAR.
• The approach offers a promising direction for developing more scalable and adaptable HAR systems.
• User-aware contrastive supervision is a key component for improving HAR model performance on unseen users.