Group Nonnegative Matrix Factorization for EEG Classification

Given electroencephalogram (EEG) data measured from several subjects under the same conditions, our goal is to estimate common task-related bases in a linear model that capture intra-subject variations as well as inter-subject variations. Such bases capture the common phenomenon in group data, which is a core of group analysis. In this paper we present a method of nonnegative matrix factorization (NMF) that is well suited to analyzing EEG data of multiple subjects. The method is referred to as group nonnegative matrix factorization (GNMF) where we seek task-related common bases reflecting both intra-subject and inter-subject variations, as well as bases involving individual characteristics. We compare GNMF with NMF and some modified NMFs, in the task of learning spectral features from EEG data. Experiments on brain computer interface (BCI) competition data indicate that GNMF improves the EEG classification performance. In addition, we also show that GNMF is useful in the task of subject-tosubject transfer where the prediction for an unseen subject is performed based on a linear model learned from different subjects in the same group.