ALGGNet: An adaptive local-global-graph representation network for brain-computer interfaces

The brain is a complex system comprising neurons, local circuits, and functional regions. Neural pathways link various regions and collaborate to accomplish intricate cognitive and behavioral functions. To investigate brain activity within and between functional regions, we propose an adaptive local-global graph representation network (ALGGNet). This network suits brain-computer interface (BCI) applications, including the decoding of motor imagery electroencephalogram (EEG) signals and emotion recognition. ALGGNet consists of three essential components: the adaptive local region partitioning block, the temporal learning block, and the graph learning block. First, we adopt the concept of brain functional regions in neuroscience and employ the double clustering algorithm to generate brain-like functional regions for specific data sets adaptively. The EEG channels are then reordered according to the clustering results . Subsequently, the EEG signal passes through the time information processing stage and the graph learning stage. The time information processing stage mainly includes a multi-scale temporal convolution block and a feature mapping block with a coordinate attention (CA) mechanism. The graph learning stage includes graph filters that aggregate information about local regions and global graph convolutional networks (GCNs) that learn complex relationships between local regions. We conduct experiments on the motor imagery dataset and the emotion recognition dataset, respectively, and the results show that the proposed network obtains state-of-the-art performance. Additionally, we perform many ablation experiments, parameter tuning experiments, and visualization experiments to demonstrate the proposed ALGGNet better.