Feature Extraction by Nonnegative Tucker Decomposition from EEG Data Including Testing and Training Observations

The under-sample classification problem is discussed for 21 normal childrenand 21 children with reading disability. We first rejected data of one subject in each group and produced 441 sub-datasets including 40 subjects in each. Regarding each sub-dataset, we extracted features through nonnegative Tucker decomposition (NTD) from event-related potentials, and used the leave-one-out paradigm for classification. Averaged accuracies over 441 sub-datasets were 77.98% (linear discriminate analysis), 73.55% (support vector machine), and 76.97% (adaptive boosting). In summary, assuming K observations with known labels, for the new observation without the group information, the feature of the new observation can be extracted through performing NTD to extract features from data of all observations (K+1). Since the group information of the first K observations is known, their features can train the classifier, and then, the trained classifier recognizes new features to determine the group information of new observation.