A Self-Adaptive Subgraph Generation Algorithm for EEG Channel Selection

Electroencephalogram (EEG) channel selection is an important issue in the fields of brain computer interfaces which can effectively reduce the amount of calculation and noise between channels. In this paper, a self-adaptive subgraph generation algorithm for EEG channel selection (SSGE), built on the base of graph convolution network (GCN), was proposed to generate a subgraph based on a learned weighted adjacency matrix. Different from the traditional GCN methods, we proposed a novel message passing mechanism to reconstruct information between nodes to facilitate better channel selection, solving the problem that traditional graph convolution can't accommodate learnable weighted adjacency matrix. We conducted experiments on the EEG three-class emotion dataset (SEED). The experiment results demonstrated that using our generated subgraph, including 17 nodes out of a total of 62 nodes and 7 edges, has achieved 82.18%, 74.83%, and 79.07% recognition accuracy for the subject independent experiment on three sessions respectively, making an average improvement over using all channels 2.53%.