Evaluation of machine learning approaches for sensor-based human activity recognition

Document Type : Research Paper


Department of Computer Science, College of Science, University of Diyala, Baqubah, Iraq


Human Activity Recognition (HAR) systems used in healthcare have attracted much attention in recent years. A HAR system consists of a wearable device with sensors. HAR has been used to suggest several machine learning (ML) algorithms. However, only a few research have looked at how to evaluate HAR to identify physical activities. Nevertheless, obtaining an explanation for their performances is complicated by two factors: the lack of implementation specifics and the lack of a baseline evaluation setup that makes comparisons unfair. For establishing effective and efficient ML–HAR of computers and networks, this study uses ten common unsupervised and supervised ML algorithms. The decision tree (DT), artificial neural network (ANN), naive Bayes (NB), k-nearest neighbor (k-NN), support vector machine (SVM), random forest (RF), and XGBoost (XGB) algorithms are among the supervised ML algorithms, while the k-means, expectation-maximization (EM), and self-organizing maps (SOM) algorithms are among the unsupervised ML algorithms. Multiple algorithms models are presented, and the turning and training parameters in ML (DT, ANN, NB, KNN, SVM, RF, XGB) of each method are investigated in order to obtain the best classifier assessment. Differ from earlier research, this research measures the true negative and positive rates, precision, accuracy, F-Score as well as recall of 81 ML-HAR models to assess their performance. Because time complexity is a significant element in HAR, the ML-HAR models training and testing time are also taken into account when evaluating their performance efficiency. The mobile health care (M\_HEALTH CARE) dataset, which includes real-world network activity, is used to test the ML-HAR models. In general, the XGB outperforms the DT-HAR, k-NN-HAR, and NB-HAR models in recognizing human activities, with recall, precision, and f-scores of 0.99, 0.99, and 0.99 for each, respectively, for health care mobile recognition.


Volume 13, Issue 2
July 2022
Pages 1183-1200
  • Receive Date: 12 January 2022
  • Revise Date: 19 February 2022
  • Accept Date: 23 March 2022
  • First Publish Date: 05 April 2022