3.4.6. Graph Neural Networks

Graph Neural Networks (GNNs) are specialized neural networks that work with data structured as graphs, where nodes represent data points, and edges define relationships or dependencies. Unlike traditional models, which process fixed data structures, GNNs let each node gather and combine information from its neighboring nodes, making them well-suited for complex, non-Euclidean data (where data points can’t be represented in straight lines or rectangular coordinates and data connections don’t follow a regular, grid-like format) such as social networks or brain connectivity patterns in EEG signals,. This structure helps GNNs capture intricate relationships in the data, though they can require high computational resources and may be sensitive to noise in the graph data.

In EEG-based emotion recognition, Gilakjani and Osman used a GNN model with contrastive learning and GAN-based data augmentation to boost performance across datasets, achieving 64.84% and 66.40% accuracy for valence and arousal classification on the DEAP dataset. Meanwhile, Jang et al. developed a GNN for EEG-based emotional video classification, which automatically learns a multi-layer graph structure to represent connectivity between brain regions, giving better performance than models that rely on predefined connectivity structures.