Community-Structure Enhanced Brain Graph Mining

Brain neuroimaging technology has become increasingly valuable in diagnosing brain disorders, with brain graphs constructed from neuroimaging data providing critical insights into the properties of the human brain. Although functional connectivity-based brain graph construction methods have proven effective in diagnostic tasks, they often fail to incorporate prior knowledge from neuroscience, such as the existence of distinct functional modules that govern various aspects of brain function. In this work, we define these functional modules as different communities of brain graphs and propose a community-structure enhanced brain graph mining framework. Specifically, we design community-constrained node identity vectors as part of node features to adaptively capture the community structure in the process of model optimization toward downstream tasks. Importantly, since common community variations are identified across subjects with the same disorder, these identity vectors are shared among all subjects within a specific brain disorder diagnostic task. Additionally, to improve the interpretability of our model, we generate attention scores solely from the identity vectors, enabling the model to constantly focus on specific brain regions and communities associated with a particular disorder. Finally, we conduct extensive experiments on three real-world datasets. The analysis results demonstrate the effectiveness of our approach and provide valuable insights for identifying biomarkers associated with corresponding brain disorders.