Comparative Analysis of Results of Modern Classification Algorithms Usage for Determining the Type of Physical Activity Based on Integrated Sensors Data

Abstract

The article considers the usage of various classification algorithms to process data from integrated sensors of a modern smartphone to obtain user's physical activity. To compare classification algorithms, we use modeling based on a dataset from an experiment that has 561 significant characteristics for each type of activity with a total number of records equal to 10299. The analysis of data from integrated sensors is performed by several classification algorithms: decision trees, Random Forest algorithm, logistic regression, support vector machine with a linear function and a kernel function. In addition to pure (without modification) ones, algorithms with preliminary optimization of hyperparameters using random search and cross-validation are used. The evaluation metrics are: Accuracy, Precision, Recall, F-measure, Mean Absolute Error and Root Mean Square Error. The estimated metrics and confusion matrices obtained using specially written python scripts using the sklearn library are provided for each algorithm. Analysis of the presented results shows that the best classification results both with preliminary optimization of hyperparameters and cross-validation, and without preliminary tuning, are provided by the use of linear support vector machine algorithm.