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

Existing explainability metrics have significant limitations: The study found that "faithfulness" — a commonly used metric to verify that AI explanations actually reflect how a model makes decisions — may not be reliable in certain contexts, such as with the EfficientGCN model, suggesting that practitioners cannot always trust standard evaluation methods.

Stability is more trustworthy than faithfulness for movement analysis: When researchers made small, realistic changes to human movement data, "stability" proved to be a more dependable metric than faithfulness, indicating that some evaluation approaches are better suited for activity recognition systems than others.

Current explanation methods lack diversity: The paper found that two popular explanation techniques (CAM and Grad-CAM) "produce almost identical explanations," which limits researchers' ability to validate whether explanations are truly capturing how the AI works or simply reflecting similar technical approaches.

Biomechanical realism matters for testing AI systems: The researchers developed a perturbation method that respects how human bodies actually move, rather than making arbitrary changes to data — highlighting that responsible AI evaluation should account for real-world constraints and human physiology.

The field needs better evaluation standards: The authors call for "more diversified metrics and new XAI methods" in skeleton-based activity recognition, indicating that current tools for verifying AI trustworthiness in this domain are insufficient and require further development.