The Impact of Feature Selection Algorithms on Emotion Recognition "in-the-wild"

Physiological sensors are increasingly integrated into wearable devices, enabling pervasive emotion recognition in-the-wild. However, processing these sensors’ data requires high computational demands, and achieving reliable emotion recognition in-the-wild remains challenging. Thus, feature selection is crucial for optimizing performance by eliminating redundant data. This paper addresses feature selection in emotion recognition in-the-wild. Therefore, we examine how various feature selection algorithms (i.e., Wilcoxon, Genetic Algorithm, Support Vector Machine, Sequential Backward Selection, Student’s T-test, Recursive Feature Elimination, and Random Forest) perform on laboratory and in-the-wild data, published by Heinisch et al. 2019. Minimal overlap in features selected was found among all 18 participants across both scenarios (i.e., laboratory and in-the-wild). The features most frequently selected were applied to build new classification models and assess their classification capabilities across both scenarios. A semi-supervised approach, with feature scaling and label masking during training, was implemented. Although this model was trained with laboratory and tested with in-the-wild data, it achieved a mean classification accuracy of 0.83 across all participants, outperforming the supervised models with a mean classification accuracy of 0.31.