Ensemble Machine-Learning Applied to Brain-Computer Interfaces
Brain-Computer Interfaces (BCI) are robust systems that incorporate techniques from different scientific fields in order to perform the translation of brain signals into external commands. Among these techniques, one can highlight machine learning algorithms, which are responsible for recognizing patterns and extracting relevant information from data to correctly perform the classification task. However, since the performance of these algorithms also depends on data complexity, the conception of a multi-user BCI system is still a challenging task, given the high intra/inter-user variability that is faced in this scenario. In this case, ensemble learning can provide promising results, considering that each base classifier acts as an expert with a subset of the dataset. Thus, in this work, we investigate the use of different ensemble structures classically disseminated in the literature to classify brain signals of different subjects. As the main contribution to this topic, we propose a new type of ensemble: the gaussian kernel-based Mixture of Experts (GKME). Experimental results shows that GKME can reach a better performance when compared to classic ensemble algorithms with a significantly lower number of experts, achieving a mean accuracy of 95.33% when evaluating subjects separately and 85.14% in the cross-subject scenario.