Brain Graph Augmentation via Learnable Edge Masking for Psychiatric Diagnosis (arxiv.org)

Researchers introduce SAM-BG, a two-stage framework that improves representation learning for brain connectivity graphs by using a learnable edge masker to preserve structural semantics during augmentation. In the first stage, an edge masker is trained on a small labeled subset to identify clinically meaningful connections; in the second, those learned priors guide self-supervised, structure-aware augmentations so the model can learn robust, semantically faithful graph embeddings without corrupting key topology. The approach is motivated by the observation that common graph augmentation strategies can inadvertently destroy diagnostically important patterns in brain networks. On two real-world psychiatric datasets, SAM-BG outperforms state-of-the-art methods—especially in low-label regimes—and produces interpretable connectivity patterns linked to clinical phenomena. Technically, the contribution is a domain-informed augmentation policy: masking decisions are learned rather than random, enabling contrastive or other SSL objectives to operate on perturbations that respect neurobiological structure. Implications include more data-efficient psychiatric diagnosis, improved model interpretability for neuroscientists and clinicians, and a template for designing semantics-preserving augmentations in other biomedical graph domains. Code is publicly available for replication and extension.