๐Ÿ“š https://doi.org/10.1038/s41598-024-83475-4

โœ… Day 3 โ€“ Alignment, Ensemble, and SHAP: Inside the Core Mechanism

Today I focused on the core algorithmic design of the proposed framework โ€” how it turns raw joint angles into explainable scores using alignment, ensemble regression, and SHAP-based interpretation.


๐Ÿง  What I Learned โ€“ From Feature to Final Score

๐Ÿ”ง Skeleton Feature Extraction

\[A_i = \arccos \left( \frac{\vec{v_1} \cdot \vec{v_2}}{|\vec{v_1}||\vec{v_2}|} \right) \cdot \frac{180}{\pi}\]

โ†’ This captures detailed movement articulation, including elbow bends, knee twists, torso rotation, etc.


๐Ÿ“ Feature Alignment


๐Ÿง  Regression Models + Adaptive Ensemble

\[w_i' = \left( e^{|\text{RMSE}_i - \text{RMSE}_{\max}|} \right)^k, \quad w_i = \frac{w_i'}{\sum_j w_j'}\] \[\hat{y}_{\text{final}} = \sum_i w_i \cdot \hat{y}_i\]

โ†’ Ensemble behaves like a human judging panel, where better performers influence more.


๐Ÿ’ก Explainability with SHAP


๐Ÿ” Insight Snapshot

โ€œSpatial/temporal alignment + explainable ensemble = evaluation system thatโ€™s both accurate and usable.โ€


๐Ÿ› ๏ธ What Iโ€™ll Build Next


๐Ÿ”ญ Day 4 Preview


๐Ÿ“ Reflection

Today clarified how much alignment and angle definition impact downstream models.
I also liked how explainability wasnโ€™t treated as an afterthought โ€” itโ€™s built into the pipeline.

Iโ€™m now seriously thinking about how to embed this structure into dance feedback systems or rehab movement scoring.
Even without deep learning, this paper shows a lot can be achieved with clear structure and interpretability.