Time-Varying Graph Signal Processing Based Cross-Subject Emotion Classification from Multi-Electrode EEG Signals

Emotion classification plays an important role in the domain of human-computer interaction (HCI). In this paper, a novel framework for learning emotion-specific brain functional connectivity from EEG signals, with blockwise time-varying graph signal processing (GSP), is proposed. Graph corresponding to the last temporal block, which captures the spatio-temporal smoothness from all of the previous blocks is considered for extracting the Laplacian-based graph spectral features. The deviation range of the eigenvalues and their ratio corresponding to low-frequency and high-frequency components are proposed in the form of a convex sum feature. This feature is further utilized to classify cross-subject based positive and negative emotions using the KNN classifier. Simulation results on the benchmark DREAMER database validate the performance of the proposed method with metrics of accuracy, sensitivity, and specificity and are comparable with the existing state-of-the-art techniques.