Learning More Significant Higher-Order Interactions Among Brain Regions

We propose a novel model capturing salient higherorder interactions among brain regions, yielding deeper insights into intrinsic higher-order connectivity while elevating downstream accuracy. Pairwise methods overlook multi-region interactions, imposing inherent constraints. Current hypergraph approaches depend on fixed or empirical hyperedges, unable to adapt to diverse patterns in fMRI data. Advanced trainable mechanisms often generate overlaps, inducing redundancy, blurring interpretations, and disregarding intra-hyperedge node salience. To mitigate these, we introduce a hypergraph framework for brain networks with three innovations: 1) a feature-driven dynamic mask generation strategy that produces sample-specific masks to learn unique higher-order relationships and enhance downstream performance; 2) a regularization term that penalizes hyperedge overlap to promote non-redundant connections and distinctive connectivity patterns; and 3) an attention mechanism for aggregating node features within hyperedges to focus on more important brain regions. Extensive experiments on ABIDE, ADHD, and HCP-100 datasets demonstrate our model's ability to effectively capture significant higher-order brain relationships and improve task accuracy.