Structure-Centric Graph Foundation Model via Geometric Bases
Researchers have introduced Structure-Centric Graph Foundation Models (SCGFM), a novel approach designed to overcome limitations in existing graph foundation models related to structural heterogeneity and incompatible node feature spaces. By treating graph topology as the primary source of transferable knowledge, SCGFM models graphs as metric measure spaces and introduces learnable geometric bases to define a shared structural coordinate system. The model aligns graphs to these bases using Gromov-Wasserstein distances, producing structure-aligned latent representations that effectively handle diverse graph topologies. Additionally, it employs a structure-aware feature re-encoding mechanism to unify node representations without requiring fixed feature dimensions or dataset-specific preprocessing. Experimental results on both graph-level and node-level tasks demonstrate that SCGFM achieves strong generalization capabilities within and across domains, significantly outperforming current state-of-the-art graph foundation model approaches. This development represents a significant advancement in machine learning techniques for handling complex, heterogeneous graph data structures.
Wire timeline
Structure-Centric Graph Foundation Model via Geometric Bases
Researchers have introduced Structure-Centric Graph Foundation Models (SCGFM), a novel approach designed to overcome limitations in existing graph foundation models related to structural heterogeneity and incompatible node feature spaces. By treating graph topology as the primary source of transferable knowledge, SCGFM models graphs as metric measure spaces and introduces learnable geometric bases to define a shared structural coordinate system. The model aligns graphs to these bases using Gromov-Wasserstein distances, producing structure-aligned latent representations that effectively handle diverse graph topologies. Additionally, it employs a structure-aware feature re-encoding mechanism to unify node representations without requiring fixed feature dimensions or dataset-specific preprocessing. Experimental results on both graph-level and node-level tasks demonstrate that SCGFM achieves strong generalization capabilities within and across domains, significantly outperforming current state-of-the-art graph foundation model approaches. This development represents a significant advancement in machine learning techniques for handling complex, heterogeneous graph data structures.
cs.AI updates on arXiv.org