[Paper Review] Choose Your Explanation: A Comparison of SHAP and GradCAM in Human Activity Recognition

Felix Tempel, Daniel Groos+3 more|arXiv (Cornell University)|Dec 20, 2024

Context-Aware Activity Recognition Systems4 citations

TL;DR

This study compares SHAP and GradCAM for explainable AI in human activity recognition using graph convolutional networks on skeleton-based data from real-world datasets, including a cerebral palsy case. It finds that SHAP offers detailed, feature-level interpretability for clinical trustworthiness, while GradCAM delivers faster, spatially localized explanations ideal for real-time use, with both methods being complementary rather than competing.

ABSTRACT

Explaining machine learning (ML) models using eXplainable AI (XAI) techniques has become essential to make them more transparent and trustworthy. This is especially important in high-stakes domains like healthcare, where understanding model decisions is critical to ensure ethical, sound, and trustworthy outcome predictions. However, users are often confused about which explanability method to choose for their specific use case. We present a comparative analysis of widely used explainability methods, Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), within the domain of human activity recognition (HAR) utilizing graph convolutional networks (GCNs). By evaluating these methods on skeleton-based data from two real-world datasets, including a healthcare-critical cerebral palsy (CP) case, this study provides vital insights into both approaches' strengths, limitations, and differences, offering a roadmap for selecting the most appropriate explanation method based on specific models and applications. We quantitatively and quantitatively compare these methods, focusing on feature importance ranking, interpretability, and model sensitivity through perturbation experiments. While SHAP provides detailed input feature attribution, Grad-CAM delivers faster, spatially oriented explanations, making both methods complementary depending on the application's requirements. Given the importance of XAI in enhancing trust and transparency in ML models, particularly in sensitive environments like healthcare, our research demonstrates how SHAP and Grad-CAM could complement each other to provide more interpretable and actionable model explanations.

Motivation & Objective

• To address the lack of systematic comparison between SHAP and GradCAM in human activity recognition (HAR) using graph neural networks.
• To evaluate the strengths and limitations of SHAP and GradCAM in explaining GCN-based HAR models on real-world datasets, including a clinical cerebral palsy dataset.
• To provide practical guidance on selecting explainability methods based on application-specific needs such as interpretability depth, runtime, and spatial focus.
• To demonstrate that SHAP and GradCAM are complementary—SHAP for 'why' (feature importance), GradCAM for 'where' (spatial attention)—enhancing model transparency in healthcare.

Proposed method

• The study employs graph convolutional networks (GCNs) trained on skeleton-based human activity data from two real-world datasets: NTU RGB+D 60 and a clinical cerebral palsy (CP) dataset.
• SHAP is applied to compute feature-level attributions by estimating Shapley values, providing per-feature importance scores across input key points and their features.
• GradCAM is implemented using gradient-weighted class activation maps to generate spatial heatmaps highlighting body parts most influential in the model’s predictions.
• A perturbation-based evaluation is conducted to assess model sensitivity and interpretability, measuring how changes in input features affect predictions.
• The methods are compared quantitatively using feature importance ranking and qualitatively through visual interpretability and spatial coherence of explanations.
• The analysis includes both quantitative metrics (e.g., feature ranking consistency) and qualitative assessment of explanation clarity and clinical relevance.

Experimental results

Research questions

• RQ1 Which explainability method—SHAP or GradCAM—provides more reliable and interpretable feature importance rankings in GCN-based HAR models?
• RQ2 How do SHAP and GradCAM differ in their spatial localization accuracy and computational efficiency when applied to skeleton-based HAR data?
• RQ3 To what extent do SHAP and GradCAM capture meaningful patterns in clinical HAR data, such as in cerebral palsy diagnosis?
• RQ4 How do feature dependencies and correlations in human motion data affect the reliability of SHAP’s feature attributions?
• RQ5 Can SHAP and GradCAM be used in tandem to provide a more comprehensive explanation of GCN decisions in healthcare applications?

Key findings

• SHAP provides detailed, per-feature attributions that are essential for understanding individual key point contributions, making it suitable for high-stakes clinical decisions where accountability is critical.
• GradCAM delivers faster, spatially coherent explanations that highlight active body regions, making it ideal for real-time applications or visual feedback in clinical assessments.
• SHAP’s performance is affected by feature dependencies, particularly in kinematically correlated human motion data, due to its default independence assumption, which may dilute feature importance in correlated inputs.
• GradCAM lacks granularity in capturing complex feature interactions across layers and cannot explain individual feature contributions, limiting its use in deep diagnostic reasoning.
• The two methods are complementary: SHAP explains 'why' the model makes a decision by quantifying feature importance, while GradCAM explains 'where' by localizing spatial attention on the body.
• In clinical HAR tasks such as cerebral palsy prediction, SHAP’s detailed feature-level explanations enhance trust and transparency, while GradCAM supports rapid visual validation of model focus areas.