Emotion recognition using multi-scale EEG features through graph convolutional attention network

Emotion recognition via electroencephalogram (EEG) signals holds significant promise across various domains, including the detection of emotions in patients with consciousness disorders, assisting in the diagnosis of depression, and assessing cognitive load. This process is critically important in the development and research of brain-computer interfaces, where precise and efficient recognition of emotions is paramount. In this work, we introduce a novel approach for emotion recognition employing multi-scale EEG features, denominated as the Dynamic Spatial-Spectral-Temporal Network (DSSTNet). DSSTNet includes three main parts, the first is spatial features extractor, which converts EEG signal into graph structure data, and uses graph convolutional network (GCN) to dynamically optimize the adjacency matrix during the training process to obtain the spatial features between the channels. Next, band attention module is composed of semi-global pooling, localized cross-band interaction and adaptive weighting, which further extracts frequency information. Finally, through the temporal features extractor, the deep temporal information is extracted by stacking several one-dimensional convolutional layers. In addition, in order to improve the performance of emotion recognition and filter valid channels, we add a ℓ_2,1-norm regularization term to the loss function to make the adjacency matrix constraint sparse. This makes it easier to preserve emotionally relevant channels and eliminate noise in irrelevant channel. Combined with the channel selection method of graph theory, a small number of optimal channels are selected. We used a self-constructed dataset TJU-EmoEEG and a publicly available SEED dataset to evaluate DSSTNet. The experimental results demonstrate that DSSTNet outperforms current state-of-the-art (SOTA) methods in emotional recognition tasks.