LSTM-enhanced multi-view dynamical emotion graph representation for EEG signal recognition

Abstract Objective and Significance: This paper proposes an LSTM-enhanced multi-view dynamic emotion graph representation model, which not only integrates the relationship between electrode channels into electroencephalogram (EEG) signal processing to extract multi-dimensional spatial topology information but also retains abundant temporal information of EEG signals. Approach: Specifically, the proposed model mainly includes two branches: a dynamic learning of multiple graph representation information branch and a branch that could learn the time-series information with memory function. First, the preprocessed EEG signals are input into these two branches, and through the former branch, multiple graph representations suitable for EEG signals can be found dynamically, so that the graph feature representations under multiple views are mined. Through the latter branch, it can be determined which information needs to be remembered and which to be forgotten, so as to obtain effective sequence information. Then the features of the two branches are fused via the mean fusion operator to obtain richer and more discriminative EEG spatiotemporal features to improve the performance of signal recognition. Main results: Finally, extensive subject-independent experiments are conducted on SEED, SEED-IV, and Database for Emotion Analysis using Physiological Signals datasets to evaluate model performance. Results reveal the proposed method could better recognize EEG emotional signals compared to other state-of-the-art methods.