Fixed Low-Rank EEG Spatial Filter Estimation for Emotion Recognition Induced by Movies

In this paper, we propose a fixed low-rank spatial filter estimation for brain computer interface (BCI) systems with an application that recognizes emotions elicited by movies. The proposed approach unifies such tasks as feature extraction, feature selection, and classification, which are often independently tackled in a "bottom-up" manner, under a regularized loss minimization problem. We explicitly derive the loss function from the conventional BCI approach and solve its minimization by optimization with a non-convex fixed low-rank constraint. For evaluation, we conducted an experiment to induce emotions by movies for dozens of young adult subjects and estimated the emotional states using the proposed method. Our results show competitive performance against conventional methods using CSP. The advantage of the proposed method is the holistic approach, which combines feature extraction, feature selection and classification. The obtained results are also plausible from the standpoint of neurophysiological interpretation.