Examining the Effect of Feature Normalization and Feature Selection for Logistic Regression Based Multimodal Stress Detection

Abstract

Automated multimodal stress detection using smartwatches and machine learning (ML) has been very popular nowadays. One of the processes in ML-based classification is preprocessing, which includes feature normalization and feature selection because it can enhance classification performance. In this study, we construct a multimodal-based stress detection system using Logistic Regression and investigate the effects of feature normalization and feature selection on performance. The experiment results show that the stress classification system with feature normalization performs better than without feature normalization. The results also show that the use of the fewest features gives the worst performance. The performance of the stress classification system increases as the number of features increases but the performance slightly declines at a particular point. The best performance was obtained when Min-Max normalization and ANOVA-based feature selection were employed with accuracy, precision, recall, and F1-measure of 0.894, 0.819, 0.859, and 0.817, respectively. This best result was achieved when 90% of the total features (378 features) were used.