Enhancing human activity recognition model performance through KMeans-based stratified data splitting

Human Activity Recognition (HAR) plays a crucial role in healthcare, sports, security, and human-computer interaction. A major challenge in HAR is the imbalance and uneven distribution of features between the training and testing datasets, leading to biased machine learning models and reduced prediction performance. This study proposes a novel approach that combines clustering of KMeans with a stratified data splitting strategy. By utilizing clusters generated from KMeans, this method ensures that both the training and testing datasets contain representative features from all clusters, improving the model's reliability and generalizability. The Kolmogorov-Smirnov test is used to assess the uniformity of the feature distribution. Experimental results demonstrate that this method significantly enhances model accuracy and performance, achieving an accuracy of 98.58%, a Recall score of 98.66%, a Precision score of 98.65%, and an F1 score of 98.65%. These findings not only improve the effectiveness of current HAR models, but also open new research avenues for optimizing feature distribution in complex, multidimensional problems.