A Novel KOSFS Feature Selection Algorithm for EEG Signals

Abstract

One major area in data stream mining that has attracted the interest of researchers is online feature selection. By removing unnecessary and duplicated information, this approach decreases the dimensional of the streaming features, so developing a feature selection algorithm on large observational data is an important problem. Various algorithms have been proposed to handle this problem but most of the approaches did not consider the implications of multivariate problem.To overcome the limitations of multivariate problem, this paper presents a novel algorithm for feature selection in electroencephalography (EEG) signals called Kernel Online Streaming Feature Selection (KOSFS) which uses the kernel based conditional dependence to define the Markov blanket to accommodate the multivariate situation. This approach provides better prediction accuracy with fewer strong related features and reduced number of features.