Classification of Electroencephalograph Data: A Hubness-aware Approach

Classification of electroencephalograph (EEG) data is the common denominator in various recognition tasks related to EEG signals. Automated recognition systems are especially useful in cases when continuous, long-term EEG is recorded and the resulting data, due to its huge amount, cannot be analyzed by human experts in depth. EEG-related recognition tasks may support medical diagnosis and they are core components of EEGcontrolled devices such as web browsers or spelling devices for paralyzed patients. Stateof-the-art solutions are based on machine learning. In this paper, we show that EEG datasets contain hubs, i.e., signals that appear as nearest neighbors of surprisingly many signals. This paper is the first to document this observation for EEG datasets. Next, we argue that the presence of hubs has to be taken into account for the classification of EEG signals, therefore, we adapt hubness-aware classifiers to EEG data. Finally, we present the results of our empirical study on a large, publicly available collection of EEG signals and show that hubness-aware classifiers outperform the state-of-the-art time-series classifier.