Sparse graphic attention LSTM for EEG emotion recognition

Abstract

In this paper, a novel multichannel EEG emotion recognition method based on sparse graphic attention long short-term memory (SGA-LSTM) is proposed. The basic idea of SGA-LSTM is to adopt graph structure modeling EEG signals to enhance the discriminative ability of EEG channels carrying more emotion information while alleviate the importance of the EEG channels carrying less emotion information. To this end, we employ two graphic branches. One branch generates global features reflecting the intrinsic relationship between EEG channels and the other generates an attention vector guiding the global features to focus on specific EEG channels. Researches on brain emotion show that different brain regions may be related to different brain functions and the contribution of each EEG channel to one specific brain function are possibly sparse such that ℓ1-norm penalty is applied. Extensive experiments are conducted on our dry electrodes EEG database and MPED database. The experimental results show that the proposed method is superior to the state-of-the-art methods.