Model-Agnostic Structural Transfer Learning for Cross-Domain Autonomous Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Activity recognition models struggle with diverse smart devices in the Internet-of-Things era.
• Cross-domain adaptation is challenging, especially across different sensor modalities and feature levels.

Purpose of the Study:

• To propose ActiLabel, a framework for cross-domain activity recognition.
• To address the limitations of current models in adapting to new devices and sensor types.

Main Methods:

• Developed ActiLabel, a combinatorial framework using dependency graphs to model structural similarities.
• Learned optimal mappings between source and target domains at a structural level.
• Utilized graph models to abstract activity patterns from low-level signals.

Main Results:

• ActiLabel demonstrated superior performance compared to state-of-the-art transfer learning and deep learning methods.
• Achieved significant average F1-score improvements: 36.3% (cross-modality), 32.7% (cross-location), and 9.1% (cross-subject).
• Validated through extensive experiments on three large wearable sensor datasets.

Conclusions:

• ActiLabel effectively overcomes cross-domain adaptation challenges in activity recognition.
• The dependency graph approach provides a robust method for learning structural similarities across diverse sensor data.
• ActiLabel offers a promising solution for robust and adaptable activity recognition systems.