Random Subset Multi-domain Feature Extraction for Attentional State Recognition

Existing attentional state recognition methods achieve good results by utilizing frequency domain features, but spatial information has not been fully considered. In this paper, a random subset multi-domain feature extraction method is proposed. To exploit the spatial information, the training data is first divided into several non-overlapping subsets, and independent Riemannian manifolds are constructed within each subset. Riemannian distances from the Riemannian means are extracted as the feature. Besides, frequency domain information is extracted using a filter bank while phase domain information is extracted using a Hilbert transform. Finally, Riemannian distances from Riemannian means are extracted from multi-domain EEG signals. The influence of different filter banks and various numbers of random subsets are validated in the experiments. The idea of the random subset is implemented in the minimum distance to the Riemannian mean method and the results show its effectiveness. The proposed method achieves superior results compared with existing methods in attentional state recognition with an accuracy of 92.25 ± 4.58%.