Klasifikasi Aktivitas Manusia Menggunakan Extreme Learning Machine dan Seleksi Fitur Information Gain

Abstract

Human activity recognition has various benefits in daily lives. However, research in this area is still facing problems that is, unobtrusive data gathering, high dimensionality features, and the algorithm used to classify human activities. Those problems could impact in the result of the developed model. This research is a preliminary study in human activity recognition. Five common human activity will be recognized that is, walking, walking upstairs, walking downstairs, sitting, and standing. The dataset used in this study consist of 1500 data rows and 561 features. Feature selection is performed prior to the modeling step. Information Gain is used as the feature selection in which percentile method is used to subset the number of features in the dataset. The features are then normalized and will classified using ELM. Number of optimal hidden neuron will be searched to yield high predictive accuracy. The results show 240 feature subsets return the higher accuracy. A number of 100 hidden neuron results in highest predictive classification of human activity recognition. The classification results yield accuracy, precision, recall, and F1-score of 0.85.