Detection and Validation of Macro-Activities in Human Inertial Signals Using Graph Link Prediction.

Area of Science:

• Computer Science
• Machine Learning
• Graph Theory

Background:

• Sensor-based human activity recognition (HAR) is advancing with wearable devices.
• Current HAR methods often focus on simple, short activities using sliding windows.
• Recognizing complex macro-activities and their underlying micro-activity interactions remains a challenge.

Purpose of the Study:

• To address the gap in understanding micro-activity interactions within complex macro-activities.
• To apply graph link prediction and graph neural networks (GNNs) for automated analysis.
• To validate the effectiveness of this approach on real-world datasets.

Main Methods:

• Transforming sequences of micro-activities into micro-activity graphs.
• Processing these graphs using graph neural networks (GNNs).
• Utilizing graph link prediction to identify relationships between micro-activities.

Main Results:

• Accurate identification of interactions between micro-activities.
• Precise validation of composite macro-activities using learned graph embeddings.
• Demonstrated benefit of positional encodings in GNNs for sequence recognition.

Conclusions:

• Graph link prediction is effective for analyzing complex human activities from sensor data.
• GNNs, enhanced with positional encodings, show promise for sequence recognition in HAR.
• This approach advances the automated understanding of multi-step activities.