Stacked Generalization with Wrapper-Based Feature Selection for Human Activity Recognition

Human Activity Recognition has widespread usage in the fields of healthcare and human-centric computing, which is why it is important to build efficient and robust systems for accurate predictions for the same. Ensemble-based methods are also fast gaining acceptance for their ability to significantly enhance prediction quality and accuracy while also maintaining efficiency. In this context a stacked ensemble for predicting human activity as measured by a smartphone is described. Boruta, a wrapper-based all-relevant feature selection method is used before model training, and its effect on model metrics with filter-based methods and a hybrid of both methods compared. Stacking with Bonita gave an overall accuracy of 97.01%, which is an improvement over previous work (including improved accuracy in individual activities as well) and also better than simple variance-based filtering and the hybrid of both methods, which gave an accuracy of 94.07% and 93.43% respectively.