Subject dependent EEG emotion recognition using adaptive window selection and a parameter efficient graph convolutional capsule network
TL;DR
Analysis of the proposed EDLC-Net architecture indicates that the proposed architecture maintains relatively stable performance across different network depths and that the adaptive window selection strategy is particularly beneficial when only a limited proportion of windows is retained under the considered experimental settings.
Abstract
In EEG emotion recognition, feature extraction remains challenging due to the nonlinear and non-stationary nature of EEG signals. Therefore, this paper proposes EDLC-Net, a parameter-efficient graph convolutional capsule network for subject-dependent EEG emotion recognition. The proposed model integrates an edge-enhanced graph attention network with a capsule-based classification architecture. By introducing a dynamic routing mechanism, EDLC-Net selectively aggregates neighborhood information across EEG channels, which may help reduce the performance degradation associated with over-smoothing in deep graph convolutional networks. To reduce the parameter complexity of capsule networks, a weight-sharing strategy is adopted, decreasing the number of voting parameters while preserving classification performance. In addition, an improved Cumulative Sum (CUSUM) algorithm is introduced to dynamically identify signal segments with pronounced emotional variations. Experimental results on the public SEED and DEAP datasets show that EDLC-Net achieves competitive performance under the considered subject-dependent evaluation settings. Additional trial-level and cross-session evaluations on SEED yield accuracies of 91.55% and 85.76%, respectively. Further analyses indicate that the proposed architecture maintains relatively stable performance across different network depths and that the adaptive window selection strategy is particularly beneficial when only a limited proportion of windows is retained under the considered experimental settings.