Electroencephalography (EEG) based affective computing is a new research field that aims to find neural correlates between human emotions and the registered EEG signals. Typically, emo- tion recognition systems are personalized, i.e. the discrimination models are subject-dependent. Building subject-independent models is a harder problem due to the high EEG variability be- tween individuals. In this paper we propose a unified system for efficient discrimination of positive and negative emotions in a group of 26 users. The users were exposed to high arousal affective images and the recorded brain signals differentiated according to their positive and negative valence. Major challenge in building subject independent affective models is to iden- tify the most discriminative features between subjects. The focus of the present study is to find a relevant feature selection approach that extracts features suitable for neurophysiological interpretation and validation. Spatial (channels) and temporal (brain waves peaks and their respective latencies) features are extracted from the EEG signals. The feature selection strate- gies explored (Independent spatial and temporal feature selection, Sequential Feature Selection, Feature Elimination based on data descriptive statistics) are consistent in selecting parietal and occipital channels and late waves (P200, P300) as better encoder of the emotion valence state and less variable across subjects. These results are in line with neurophysiological hypothesis of visually elicited human emotions - brain activity correlation. The relevance of the selected features was validated by five standard and one majority vote classifiers.
展开▼
