A Novel Algorithm Based on Decision Trees in Multiclass Classification

Classification is the most important part in Brain-Computer Interface problems, where our task is to decipher the individual's (usually people with physical or verbal disorders) intention from several candidates. In our study, the MEG signals were recorded from an individual when he was shown 5 different types of video clips while our task was to process the MEG signals in each experiment to guess the type of the movie from 5 candidates. In this study, we applied various approaches to this multiclass classification problem and in the end, we proposed a novel algorithm which can also be applied to any multiclass classification problem. Suppose that we are using a decision tree and at each node, the classes are going to be divided into two groups of classes. In the proposed algorithm, we defined a criterion to find the best partitioning by using the results of only ((n)(2)) classifications between each pair of classes using training data. As a result, the algorithm is polynomial and can be applied to any multiclass problem. Moreover, as a matter of accuracy, it led us to the best accuracy (61.4%) in comparison to other routine methods. Thus, this algorithm might be a powerful tool in any multiclass classification problem.