A Comparative Study of Machine Learning Techniques for Human Activity Recognition

Abstract

Current study is based on the comparative analysis of machine learning techniques for Human Activity Recognition (HAR) to explore their performance measures and computational complexity. In our experimentation, four algorithms were handled prominently out of the clusters namely: Support Vector Machines, Random Forests, Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. These algorithms were applied using an inclusive dataset of accelerometer and gyroscope readings. The findings clearly illustrate that CNNs possessed the highest accuracy of 91%. The LSTM neural network trailed by very close 90%. The Random Forest and SVM algorithms got an accuracy of 88% and 85% respectively. Based on testing, the precision measures for CNNs and LSTM were close to one. 92 and 0. The logistic regression model achieved the highest prediction accuracy score with an accuracy score of 0.91 that is, the randomly forest model got the 2nd choke at an accuracy score of 0.00 which is followed by random Fores at an accuracy score of precisely. 6.7K and svm at 9. 87. Firstly off, the CNNs and the LSTMs showed up the highest recall scores of 0. 90 and 0. While NaviNet and SNNA achieve an accuracy score of 89, respectively, Random Forests comes at the third place with a score of 0. 87 and SVM at 0 cd. 84. And the same pattern was repeated within F1-scores where CNNs and LSTM obtained a higher value for the scores. Computational intricacy analysis brought forward the lifespan fact that SVM took the least amount of time for training as well as for prediction. And, it was proceeded by Random Forests, CNNs, and LSTM respectively. Conclusions from this work give the needed base for the further improvement of HAR systems, and CNNs appear the best algorithms to do this.