Item

Abstract

Brain networks have emerged as a crucial tool for exploring brain organization and aiding in brain disorder classification. Recent graph neural network (GNN)-based methods have demonstrated strong potential for brain network analysis. However, most of these methods insufficiently model lateralized heterogeneity induced by hemispheric asymmetry. Additionally, traditional message-passing schemes mainly rely on direct neighbors, which restricts their ability to capture high-order synergistic interactions among distant brain regions. To overcome these challenges, this paper proposes a Lateralization-Aware Multi-View High-Order Graph Learning (LMHGL) framework for brain disorder classification, which provides a unified scheme to model hemispheric asymmetry, supporting high-order intra-hemispheric interactions and multi-view collaboration at the decision level. Specifically, LMHGL introduces an anatomically constrained heat diffusion mechanism to model high-order intra-hemispheric interactions. By propagating information along structural connectivity, it captures biologically plausible multi-hop dependencies and enhances the representation of lateralized brain organization. Furthermore, a multi-view collaborative decision mechanism integrates global and hemispheric-specific features, enabling the joint characterization of holistic and lateralized pathological patterns. Extensive experiments on three real-world brain disorder datasets demonstrate that LMHGL consistently outperforms state-of-the-art methods, highlighting the effectiveness of jointly modeling hemispheric asymmetry and anatomically constrained high-order interactions for learning discriminative brain network representations.