In order to improve classification accuracy, the regularized logistic regression is used to classify single-trial electroencephalogram (EEG). A novel approach, named local sparse logistic regression (LSLR), is proposed. The LSLR integrates the locality preserving projection regularization term into the framework of sparse logistic regression. It tries to maintain the neighborhood information of original feature space, and, meanwhile, keeps sparsity. The bound optimization algorithm and component-wise update are used to compute the weight vector in the training data, thus overcoming the disadvantage of the Newton-Raphson method and iterative re-weighted least squares (IRLS). The classification accuracy of 80% is achieved using ten-fold cross-validation in the self-paced finger tapping data set. The results of LSLR are compared with SLR, showing the effectiveness of the proposed method.
Harremoes obtained some new maximal inequalities for non-negative martingales. In this paper, we get some new maximal and minimal inequalities for non-negative demimartin- gales which generalize the results of Harremoes. We also obtain an inequality for non-negative demimartingales which generalizes the result of Iksanov and Marynych. Finally we obtain a strong law of large numbers, strong growth rate and integrability of supremum for demimartin- gales which generalize and improve the result of Chow.