Electroencephalogram Emotion Recognition and Analysis Based on Feature Fusion And Stable Learning

Abstract

Background and Objective: Existing methods based on Graph Convolutional Networks (GCN) have been proven effective for electroencephalogram (EEG) Emotion Recognition. However, fully utilizing EEG electrode topology and extracting fundamental features are still areas of active exploration.Methods: We propose a Stable Fusion Graph Convolutional Network (SFGCN), a high-performance and stable classification model that integrates multiple graph topologies and stable learning. Specifically, we utilize both linear and Granger causality-based graph topology construction methods and introduce a Squeeze-and-Excitation (SE) block to enhance feature fusion, allowing for comprehensive capture of complex graph information. To address potential spurious correlations when extracting features from deep models, we introduce stable learning to ensure feature independence, thereby enhancing the performance of the model. In addition, to reduce computational complexity and data dimensions, we use causal brain network analysis to select critical channels. The Granger causality of EEG data reflects information flow and significant differences in brain functional connectivity, which facilitates the analysis of functional connectivity patterns in brain networks.Results: SFGCN achieves high classification accuracy on DEAP and DREAMER datasets, reaching about 98% and 92% respectively. On DEAP, we selected 20 key channels from 32 channels through causal brain network analysis, and demonstrated the effectiveness of these channels through comparative experiments.Conclusions: SFGCN significantly improves the accuracy and exhibits better performance and stability. The use of graph topology and the extraction of fundamental features underscore the practical application potential of this method for EEG-based emotion recognition.

Keywords: Electroencephalogram (EEG), Graph convolutional network, Feature fusion, Stable learning, Causal brain network analysis