Abstract
Feature selection algorithms are an essential part in many areas of pervasive computing such as Human Activity Recognition (HAR). Due to the increasing amount of information available and the large size of the HAR datasets, it has become important to reduce the dimensions of the data to minimize the computational costs of the recognition systems. This is particularly important for devices with limited computational and battery capacities such as smartphones. In this paper, we present FeSNOC, a novel feature selection algorithm based on the ecological measure of niche overlap. The niche overlapping coefficient is used as a measure of the similarity between classes and hence as an indicator of the classification accuracy. FeSNOC is a hybrid algorithm that combines the accuracy (obtained by using the K-Nearest Neighbor algorithm) and the overlapping coefficient to find the best features in a dataset. This algorithm consists of four simple steps, which do not require large computational times. We evaluate FeSNOC by using one publicly available dataset containing 561 features. We show that FeSNOC successfully selects the best features for the HAR problem outperforming other traditional feature selection approaches, such as the Information Gain and the Relief-F.