GRASP: Graph Augmentation via Sampling and Permutation

Ahmed Nebli, Helia Sahebghadam

Keywords: Structural brain graphs, graph augmentation, topology preservation

TL;DR: A brain graph augmentation technique that preserves graph topology without using deep learning

Abstract: Structural brain graphs illuminate individual differences and neurological traits but are underutilized due to limited data from the challenges of MRI acquisition and preprocessing. We introduce Graph Augmentation via Sampling and Permutation (GRASP), a method that synthesizes brain graphs by sampling edge values from consistent positions across multiple adjacency matrices within the same class---assuming topological consistency. Unlike deep learning techniques, GRASP relies on straightforward manipulations of adjacency matrices, which reduces computational demands and simplifies implementation. In this paper, we examine the proof of concept of this augmentation technique on a gender classification task using structural connectomes. We demonstrate enhanced brain graph classification and confirm that within-class adjacency consistency can generate graph variants without complex modeling. The code is publicly available at: https://github.com/heliasah/GRASP-Code

Submission Number: 64