From Movements to Metrics: Evaluating Explainable AI Methods in Skeleton-Based Human Activity Recognition.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep learning for human activity recognition (HAR) using 3D skeleton data is vital for healthcare, security, and HCI.
• A gap exists in evaluating the reliability of Explainable AI (XAI) metrics within skeleton-based HAR.

Purpose of the Study:

• To assess the applicability and reliability of established XAI metrics (faithfulness, stability) on Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM) for skeleton-based HAR.
• To introduce a novel perturbation method for testing XAI metric robustness against sensor noise within human movement error tolerances.

Main Methods:

• Utilized the NTU RGB+D 60 dataset and the EfficientGCN architecture for HAR model training and testing.
• Applied a systematic perturbation method to 3D skeleton data, introducing controlled displacements to evaluate XAI metric performance.
• Assessed faithfulness and stability metrics on CAM and Grad-CAM explanations across multiple action classes.

Main Results:

• Faithfulness demonstrated inconsistent reliability across action classes for the EfficientGCN model.
• Stability emerged as a more robust and dependable XAI metric across varying perturbation magnitudes.
• CAM and Grad-CAM produced highly similar explanations and metric outcomes.

Conclusions:

• Faithfulness may have limited applicability as a sole XAI evaluation metric in skeleton-based HAR.
• Stability is a more dependable metric for assessing XAI in this domain.
• Further research into diverse XAI methods and metrics is necessary for comprehensive evaluation in skeleton-based HAR.