Optimization of Daily Physical Activity Recognition with Feature Selection

Daily physical activity monitoring and recognition have become a big health caring challenge in modern times. Fast recognition of physical activity from wearable sensors dataset with acceptable accuracy has got great research attention. In this paper, we have presented an optimization framework with feature selection techniques. The Bayesian Optimization algorithm has been employed to optimize hyper-parameters of Support Vector Machine(SVM), Random Forest (RF), Extreme Gradient Boosting(XGBoost). Two feature selection algorithms like Fast Correlation Based Filter (FCBF) and Maximum Redundancy Maximum Relevance(mRMR) feature selection have been applied to reduce the size of the extracted feature vector. Classification performances of the two feature selection techniques are compared in terms of accuracy and F1 score. For reduced feature vector the highest acceptable accuracy and F1 score of 95.4% and 94.76% respectively have been achieved by optimized XGBoost classifier with mRMR feature selection. In addition to this, it is shown that reduced computation time with considerable classification performance can be achieved by an efficient feature selection algorithm which is useful for designing a simplified system.