Comparative Study of Different Ensemble Compositions in EEG Signal Classification Problem

Abstract

The leading perspective of this paper is an introduction to three $\left( Type-I, Type-II,\,and\,Type-III\right)$ types of ensemble architectures in Electroencephalogram (EEG) signal classification problem. Motor imagery EEG signal is filtered and subsequently used for three different types of feature extraction techniques: Wavelet-based Energy and Entropy $\left( EngEnt\right)$, Bandpower $\left( BP\right)$, and Adaptive Autoregressive (AAR). Ensemble architectures have been used in various compositions with different classifiers as base learners along with majority voting as the combined method. This standard procedure is also compared with the mean accuracy method obtained from multiple base classifiers. The Type-I ensemble architecture with EngEnt and BP feature sets provides most consistent performance for both majority voting and mean accuracy combining techniques. Similarly, Type-II architecture with EngEnt and AAR feature sets provides most consistent performance for both majority voting and mean accuracy combining techniques. However, the Type-III ensemble architecture contributes highest result $82.86\%$ with K-Nearest Neighbor $\left( KNN\right)$ classifier among all three types.