Table 1.

Accuracy of character prediction in percentage

	Test set (20)	Ind. test set (100)
Our Algorithm 1 with channels Fz, Cz and Pz (5 training characters)	82.5	83.5
Our Algorithm 3 with time segment 200–500 ms (5 training characters)	90	88.5
Our Algorithm 2 (5 training characters)	92.5	93.5
Semi-supervised SVM without feature selection (5 training characters)	87.5	85.5
Standard SVM (5 training characters)	77.5	80
Rakoto’s method (85 training characters)	Not applicable	96.5
SWLDA (5 training characters)	85	84.5

The results of both our Algorithm 2 and the standard SVM are based on the training set with 5 characters, while Rakoto’s results (Rakotomamonjy and Guigue 2008) were based on 85 characters as training data. “Test set” indicates the 20 unlabelled characters used in retraining in Algorithm 2. “Ind. test set” indicates the independent test set with 100 unlabeled characters, which is the same as the test data set in the Competition

The bold values indicate best performance