<![CDATA[The increasing adoption of smartphone-based Human Activity Recognition (HAR) systems has led to the generation of high-dimensional sensor features that may introduce redundancy, increase computational complexity, and reduce model efficiency. Although feature selection techniques have been widely investigated, limited studies have explored explainable artificial intelligence approaches for progressive sensor feature minimization while maintaining classification stability. This study proposes a SHAP-guided framework for sensor feature minimization in HAR using the UCI Human Activity Recognition Using Smartphones dataset containing 561 extracted features from accelerometer and gyroscope signals across six human activities. The study employed a quantitative experimental approach using a Random Forest classifier combined with SHAP for feature importance analysis and ranking. Progressive feature reduction experiments were conducted using subsets of 300, 200, 100, 50, and 25 features. The results demonstrated that reducing the feature set from 561 to 100 features achieved approximately 82.17% feature reduction while maintaining competitive classification performance with only a minor decrease in accuracy from 92.50% to 91.75%. Furthermore, the SHAP-guided approach produced lower standard deviation values compared with random feature selection, indicating improved stability and reproducibility across repeated experiments. The novelty of this research lies in the integration of SHAP-based explainability with progressive sensor feature minimization and stability analysis in HAR, providing an interpretable and systematic framework for reducing sensor dimensionality while preserving reliable classification performance.]]>
Copyrights © 2026
