부산대학교 도서관

The paper presents a comparative discussion of classification approaches for human activity recognition tasks based on the feature sets through extensive feature selection techniques. The original dataset on Human Activity Recognition from Continuous Ambient Sensor is collected from UCI machine learning repository. Amongst the 12 activities mentioned in the dataset, five specific activities (Watching TV, Reading, Talking over Phone, Cooking and Eating) have been selected for the purpose of this research. The scraped dataset is analyzed through four feature selection methods for extracting important features upon statistical significance of features and node impurity. From the actual dataset with 37 attributes, the feature selection methodologies give four distinct features sets. Later on, Principal Component Analysis is applied on the five feature sets including the original scraped dataset to reduce feature space and five principal components are selected to cover more than 90% data variance of the feature sets. Based on the 37 features present in the actual dataset and obtained sets of important features, performance of five classifier models (K Nearest Neighbors, Decision Tree, Random Forest, Gaussian Naïve Bayes and MLP classifier using Back propagation) are evaluated. The selection of feature set based on different approaches of feature importance generates difference in outputs for each feature set on each classifier. The result shows that Multi-layer Perceptron using Back propagation algorithm achieves better accuracy on human activity recognition on the five feature sets. The research findings highlight the necessity of data preprocessing and significant feature selection for getting better accuracy score for noisy time-series data of HAR activity.