A novel data mining scheme for smartphone activity recognition by accelerometer sensor

Abstract

The prime objective of activity recognition is to recognize the actions performed by a person with the surrounding environment and forming different observation sets. It is necessary to choose the appropriate classifier for the data collected through accelerometer sensors incorporated in mobile phones, which have limited resources such as energy and computing power. In this paper, standard classification techniques of data mining like random forest (RF), multilayer perceptron (MLP), logistic regression, classification via regression, and J48 and RepTree have been implemented to compare the performance and accuracy of different classifiers by reducing the computational cost. In this experiment, it was found that RF required quite short time than MLP (0.64 vs. 270.07 s, respectively) to build the model and gives the better accuracy (92.6% vs. 92.1%, respectively). This study has concluded that RF has better performance score than other classification techniques applied in this study.