Pet Behavior Classification based on Graph Attention Network

Companion animals cannot speak, so interpreting the sensor is the diagnosis. This proposes classifying daily activity from accelerometer and gyroscope signal collected by wearables.

• An LSTM encoder extracts features from the sensor data, and those become the input to a GAT
• Static and dynamic behaviors are distinguished together
• The goal is to tell the guardian about a drop in activity sooner

This is the basis for my master’s thesis, Design and Implementation of a Graph Neural Network-based Behavior Classification Model — 91.1% test accuracy across 5 behaviors.