Explaining Human Activity Recognition with SHAP: Validating insights with perturbation and quantitative measures.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Human Activity Recognition (HAR) is crucial for high-risk applications.
• Graph Convolutional Networks (GCNs) are effective for HAR using skeleton data.
• Interpreting GCN decisions in HAR remains a challenge.

Purpose of the Study:

• To explain GCN decision-making in HAR using SHapley Additive exPlanations (SHAP).
• To evaluate the influence of specific body key points on HAR model predictions.
• To enhance model trustworthiness in critical applications like healthcare.

Main Methods:

• Applied SHAP to explain GCNs on cerebral palsy (CP) classification and NTU RGB+D 60 datasets.
• Introduced a novel perturbation approach to assess body key point influence.
• Used informed perturbation against random perturbation to validate SHAP explanations.

Main Results:

• SHAP-identified important body key points significantly impacted HAR model accuracy, specificity, and sensitivity.
• Perturbation analysis confirmed the influence of key body points on prediction outcomes.
• SHAP provided granular insights into feature contributions for GCNs in HAR.

Conclusions:

• SHAP effectively explains GCNs in HAR tasks, highlighting influential body key points.
• The findings support the development of more interpretable and trustworthy HAR models.
• This research has implications for high-stakes applications requiring reliable activity recognition.