Feature Selection and Combination for Stress Identification Using Correlation and Diversity

Using multiple physiological sensors to detect different stress level has become an important and popular task in improving human health and well-being. In the process, the selection of a smaller set of independent features is a necessary, yet challenging, step for feature combination, situation analysis and decision making. In this paper, we investigate feature selection methods using both concepts of correlation and diversity. Six feature combination methods (C4.5, Naïve Bayes, Linear Discriminant Function, Support Vector Machine, K Nearest Neighbors and Combinatorial Fusion) are applied to the selected features in the detection of the stress levels. Our results demonstrated that (a) diversity based feature selection is as good as correlation based selection across all six combination methods, and (b) combinatorial fusion method performs better than five other combination methods across all features selected by using both correlation and diversity.