A Semisupervised Recurrent Convolutional Attention Model for Human Activity Recognition.

Area of Science:

• Machine Learning
• Wearable Sensor Technology
• Human Activity Recognition

Background:

• Deep learning has advanced human activity recognition (HAR).
• Semisupervised learning is crucial due to limited labeled data in HAR.
• Class imbalance in labeled data is a significant, yet underexplored, challenge in HAR.

Purpose of the Study:

• To propose a semisupervised deep model for imbalanced activity recognition from multimodal wearable sensory data.
• To simultaneously address multimodal sensor data challenges (variability, similarity) and data limitations (scarcity, imbalance).
• To develop a robust framework for accurate HAR even with minimal labeled and imbalanced datasets.

Main Methods:

• A pattern-balanced semisupervised framework to extract and preserve diverse latent activity patterns.
• Recurrent convolutional attention networks to exploit multi-modal sensor independence and identify salient activity regions.
• Utilizing multimodal wearable sensory data for enhanced HAR.

Main Results:

• The proposed model achieved competitive performance against state-of-the-art semisupervised and supervised methods.
• Demonstrated strong results with only 10% labeled training data.
• Showcased robustness and effectiveness on imbalanced and small training datasets.

Conclusions:

• The developed semisupervised deep model effectively handles imbalanced data and limited labels in HAR.
• The pattern-balanced framework and attention mechanisms contribute to robust and accurate activity recognition.
• This approach offers a promising solution for real-world HAR applications with data constraints.