SAGN: Sparse Adaptive Gated Graph Neural Network With Graph Regularization for Identifying Dual-View Brain Networks.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Network Science

Background:

• Brain network analysis is challenged by threshold selection, leading to topological degradation and noise.
• Existing methods often overlook weakly coupled connections, limiting the understanding of complex brain systems.
• Graph Neural Networks (GNNs) show limitations in robustness and overfitting for brain network identification.

Purpose of the Study:

• To explore the untapped potential of weakly coupled brain networks in emotion identification.
• To develop a robust GNN model capable of integrating information from both strongly and weakly coupled brain networks.
• To validate the proposed model's performance using both public and custom datasets.

Main Methods:

• Pioneering construction and validation of weakly coupled brain networks for emotion identification.
• Proposal of a Sparse Adaptive Gated Graph Neural Network (SAGN) integrating dual-view brain network topologies (strong and weak).
• Implementation of a sparse adaptive global receptive field, a gated mechanism for feature enhancement and noise suppression, and graph regularization for improved generalization.

Main Results:

• Weakly coupled brain networks contain separable physiological patterns essential for distinguishing emotional states.
• The proposed SAGN model demonstrates superior performance in emotion identification tasks.
• Experiments confirm the value of weakly coupled connections and the robustness of the SAGN.

Conclusions:

• Weakly coupled brain networks hold significant, previously untapped, value for understanding brain function, particularly in emotion recognition.
• The developed SAGN model offers a robust and generalizable approach for brain network analysis by integrating diverse connection strengths.
• This research opens new avenues for utilizing complex brain network dynamics in machine learning applications.