MEG Mind Reading: Strategies for Feature Selection

The regularized logistic regression classifier has shown good performance in problems where feature selection is critical, including our recent winning submissions to the ICANN2011 MEG mind reading challenge [Huttunen et al. 2011; Huttunen et al. 2012], and to the DREAM 6 AML classification challenge [Manninen et al. 2011]. The benefit of the method is that it includes an embedded feature selection step, which automatically selects a good subset of input features, thus, simplifying the classifier and improving the generalization. However, explicit wrapper feature selection methods, such as the forward and backward feature selection, are also widely used in similar problems. In this paper, we compare the efficiency of the elastic net regularized logistic regression classifier with the support vector machine classifier in combination with various sequential feature selection methods.