📚 https://doi.org/10.1038/s41598-024-83475-4

✅ Day 4 – Experiments, Performance, and Interpretability in Practice

Today I explored how the proposed system performs across different datasets, how it compares with human experts, and how SHAP helps make the scoring explainable and educational.

📊 What I Learned – From Metrics to Meaning

🧪 Evaluation Metrics

• Used MAE, RMSE, sMAPE, R², Pearson correlation, and ICC
• LOOCV (Leave-One-Out Cross Validation) used due to small sample size
• These metrics cover both accuracy and consistency with expert scores

🔬 Performance Across Models

• Feature alignment greatly improved performance across all models
• The proposed adaptive weighted ensemble achieved the best scores
• On XSQ and PBB datasets, model even outperformed human experts in MAE
• TaiChi results still favored expert judges — possibly due to subtler motion nuance

👥 Comparison with Experts

→ Indicates strong generalization in most styles, but expert intuition still matters in complex domains.

💡 SHAP in Action

• Global SHAP: Key angles in frames 13–17, 19–23, and 31 had the biggest influence
• Local SHAP: Showed which joints in which frames led to deductions or boosts
• Partial dependency plots revealed optimal angle ranges per joint
• Feedback is interpretable and coachable — not just numerical

🔍 Insight Snapshot

“The model doesn’t just score — it teaches.”

• Even basic regressors perform well with good features and alignment
• SHAP transforms black-box regression into personalized feedback
• Ensemble + alignment bridges statistical accuracy with human-level usability

🛠️ What I’ll Build Next

• Compare SHAP output with actual dance instructor annotations
• Build a SHAP-driven feedback system to flag common joint issues
• Expand pipeline to support multi-style templates (e.g., waacking vs. hip-hop)
• Explore real-time scoring using simplified keypoint subsets

📝 Reflection

This part of the paper really convinced me that alignment and explanation together can make ML feedback systems actually useful — not just technically accurate.

It also gave me inspiration for using this approach in rehabilitation or education, where transparency matters more than raw accuracy.