Explaining Human Activity Recognition with SHAP: Validating Insights with Perturbation and Quantitative Measures

2024-11-06

Comput. Biol. Medicine

AI Summary

To address the insufficient interpretability of skeleton-based human activity recognition (HAR) models, this paper proposes the first interpretability validation framework integrating SHAP attribution analysis, controllable temporal perturbation, and multi-dimensional fidelity quantification. Methodologically, it couples sliding-window CNN/LSTM architectures with skeleton-specific joint perturbation strategies and systematically evaluates explanation reliability using deletion/insertion curves and Spearman rank correlation. The key contribution lies in the first deep integration of SHAP explanations with dynamic perturbation experiments and quantitative validation—overcoming the limitations of conventional qualitative visualizations. On benchmarks including UCI HAR, SHAP-based explanations achieve an average ΔAUC < 0.08, significantly outperforming LIME and Grad-CAM. This establishes a verifiable interpretability paradigm for trustworthy HAR model deployment in high-stakes applications.

Technology Category

• Humans and AI: Explainable AI (XAI) for Human Understanding
• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Problem

• Explaining GCN decisions in HAR using SHAP for interpretability
• Validating SHAP insights via perturbation on key body points
• Assessing impact of influential features on model performance metrics

Innovation

• Uses SHAP to explain GCN decisions
• Introduces novel perturbation for validation
• Targets key body points for fidelity