An exploration of EEG-based, non-stationary emotion classification for affective computing
University of New Brunswick
The monitoring of emotional state is important in the prevention and management of mental health problems and is increasingly being used to support affective computing. Researchers are exploring various modalities from which emotion can be inferred, such as through facial images or via electroencephalography (EEG) signals. Current research commonly investigates the performance of machine-learning-based emotion recognition systems by exposing users to short films that are assumed to elicit a single known emotional response. Assuming static emotions, even for these brief periods, however, does not consider that emotions evolve. Moreover, in order to demonstrate better results, many existing models are not tested in ways that reflect realistic real-world implementations. In this thesis, the dynamic evolution of emotions induced using longer and variable stimuli is explored using EEG signals from the publicly available dataset, AMIGOS. A variety of feature engineering and selection techniques are applied and evaluated across four different cross-validation frameworks. The role of imperfect labelling of ground truth emotions and both data and gender-imbalances in the dataset are also investigated. Improved feature design and selection lead to up to 13% absolute improvement relative to comparable previously reported studies using this dataset. Alternative training configurations and a selective confidence-based classification scheme are proposed, leading to further possible improvements.