A Novel Algorithm for Human Activity Determination Feature Selection and Classification

Abstract

Human Activity Determination (HAD) used the integrated sensors in a mobile phone is a very active research field to predict the everyday activities of humans in Machine Learning (ML). However, there are numerous challenges to develop systematically in the HAD system. In ML, feature selection is a critical issue due to containing redundant or irrelevant features to represent the target activity. Because of the large size of the dataset and the complexity of the features concerned, HAD relies mainly on feature selection to improve robustness and precision. The primary objective of this research is to discover the best ideal classifier model-based feature selection technique for HAD and Machine Learning (ML) problems. As the aim of this, an Artificial Neural Network (ANN) model using Multi-Layer Perceptron (MLP) is designed with formulating to discover the number of hidden nodes in neuron. In addition, two novel feature selection methods: 1) LDC (Linearly Dependent Concepts) using linearly dependent concepts, and 2) CAT (Cyclic Attribution Technique) using group theory and basic properties of cyclic group are identified. Three datasets (UCI-HAR, DATASET-UCI dataset, HAPT) and five different classifiers (SVM, BAG, KNN, CART, and BAYES) support to conduct the statistical and comparative analyses. Based on all systematic experiments, the performances with running time compared with each other. Although the CAT feature selection method could reduce more 30% of features than the LDC feature selection method, the accuracy of model-based LDC is better than the CAT model. The study concluded that the MLP model-based LDC approach was the most comprehensively applicable and effective methodology for HAD systems development.