Joint Rayleigh Coefficient Maximization and Graph Based Semi-Supervised for the Classification of Motor Imagery EEG

Classifying electroencephalogram (EEG) signals is one of the most important issues on motor imagery-based Brain computer interfaces (BCIs). Typically, such classification has been performed using a small training dataset. To date, most of the classification of the algorithms were proposed for large samples. In this paper, a combination of Rayleigh coefficient maximization and graph-based method was developed to classify EEG signals with small training dataset. The Rayleigh coefficient maximization was adopted to obtain the projection directions, which extract discriminating features from the preprocessed dataset. Next, both training and testing features are applied to construct an affinity matrix, and then both affinity matrix and all label information are applied to train a classifier based on graph-based semi-supervised method. In this approach, both labeled and unlabeled samples are used for training a classifier. Hence it can be used in small training data case. Finally, a new iteration mechanism is applied to update the training data set. And the experiment results on BCI competition III dataset IVa show that the classification accuracy using our method was higher than using CSP (common spatial patteru) and support vector machine (SVM) method in all subjects with different size of training dataset. We used an eightfold cross-validation on this dataset, and the results show a good stability of our algorithm.