Application of annotation smoothing for subject-independent emotion recognition based on electroencephalogram

Abstract

In the construction of computational models to recognize emotional state, emotion reporting continuously in time is essential based on the assumption that emotional responses of a human to certain stimuli could vary over time. However, currently existing methods to annotate emotion in temporal continuous fashion are confronting various types of challenges. Therefore, the manipulation of the annotated emotion prior to labeling training samples is necessary. In this work, we present an early attempt to manipulate the emotion annotated in arousal-valence space by applying three different signal filtering techniques to smooth annotation data; moving average filter, Savitzky-Golay filter, and median filter. We conducted experiments of emotion recognition in music listening tasks employing brainwave signals recorded from an electroencephalogram (EEG). Smoothed annotation data were used to label the features extracted from EEG signals to train emotion recognizers using classification and regression techniques. Our empirical results indicated the potential of the moving average filter that could increase the performance of emotion recognition evaluated in subject-independent fashion.