Weighted informative inverse active class selection for motor imagery brain computer interface

EEG signal is used to establish a communication channel between brain and control device. A common scenario of machine learning based brain computer interface (BCI) is class wise accuracy is different from overall accuracy. Accuracy for one class is worse than overall accuracy. Because, error rates of all classes are not converging in same rate during training of learner. Active class selection (ACS) can be a solution to non-uniform converging of error rates for different classes. ACS determines the class-wise proportion of samples in training set based on empirical error. One of the ACS methods is inverse proportion. It implies that class proportion is inverse of class accuracy. Entropy of unlabeled samples is an active learning (AL) approach of finding the most informative samples. The intuition of this work is to feed the learner with most informative samples maintaining the class proportion given by inverse ACS method. Following this idea, an improved weighted inverse ACS method is developed and combined with AL query concept for BCI. This proposed weighted informative inverse active class selection method has been applied on BCI competition IV motor imagery (MI) binary class data set 2B. It shows better or similar performance on most of the subjects with less amount of training samples. So, this method gives us a way of performance improvement as well as reduction of training samples in BCI.