Generalization of Local Temporal Correlation Common Spatial Patterns Using Lp-norm (0 < p < 2)

As one of the effective feature extraction methods, common spatial patterns (CSP) is widely used for classification of multichannel electroencephalogram (EEG) signals in the motor imagery-based brain-compute interface (BCI) system. The formulation of the conventional CSP based on L2-norm, however, implies that it is sensitive to the presence of outliers. Local temporal correlation common spatial patterns (LTCCSP), as an extension of CSP by introducing the local temporal correlation information into the covariance modelling of the classical CSP algorithm, extracts more discriminative features. In order to further improve the robustness of the classification, in this paper, we generalize the LTCCSP algorithm by replacing the L2-norm with Lp-norm (0 < p < 2) in the objective function, called LTCCSP-Lp. An iterative algorithm is designed under the framework of minorization-maximization (MM) optimization algorithm to obtain the optimal spatial filters of LTCCSP-Lp. The iterative solution is justified in theory and the effectiveness of our novel proposed method is verified by experimental results on a toy example and datasets of BCI competitions.