Emotional State Classification with Distributed Random Forest, Gradient Boosting Machine and Naïve Bayes in Virtual Reality Using Wearable Electroencephalography and Inertial Sensing
Nazmi Sofian Suhaimi, James Mountstephens, Jason Teo
Abstract
Among the various neurophysiological signal devices used for emotion classification, the collection of the human brain signal using an EEG device is the most effective way of measuring since it is portable, easy to set up and it is low-cost. The EEG device can record the different rhythmic bands (Delta, Theta, Alpha, Beta, Gamma) as well as provide inertial sensing data (gyroscope and accelerometer) which was used as this study's dataset. Furthermore, this study uses virtual reality as the platform to deliver 360-video stimuli that were designed and stitched according to the Arousal-Valence Space (AVS) model which focuses on four emotions selected from each quadrant that were representative to these emotions namely happy, angry, boring and calm which encompasses the high and low arousal states and negative and positive valences. The dataset was then classified using Distributed Random Forest (DRF), Gradient Boosting Machine (GBM) and Naïve Bayes (NB). The performance of the classifiers were compared using the five rhythmic bands with and without the inertial sensing data. The study shows that in the subject-dependent approach, classification performance improved when inertial sensing data were included as additional sensor modalities to serve as input features in the dataset that was fed to the machine learning classifiers with GBM and NB obtained classification accuracy of 67.04% and 36.24% respectively, DRF achieved classification accuracy of 82.49%.