Hyperbolic Space Path Aggregation for Over-squashing in Multi-relational Graph

ICLR 2026 Conference Submission17399 Authors

Keywords: over-squashing, hyperbolic space, path aggregation

Abstract: The link prediction task has attracted significant attention from the graph communities. However, GNN-based methods still exhibit subpar performance in the link prediction task for large-scale, multi-relational knowledge graphs. Previous works utilize hyperbolic space to model hierarchical relations and employ path aggregation to alleviate the over-smoothing problem. The two approaches are complementary in their advantages, but both encounter the issue of over-squashing. The former experiences spatial flattening due to curvature collapse during training, while the latter struggles to distinguish the similar entities in long-distance paths. Specifically, we utilize hyperbolic space in the message and aggregation process to curvature stability and anti-symmetry weight in the update process to alleviate the issue of over-squashing. Our method achieves improvements on two standard transductive datasets and eight inductive versions. Further analysis reveals the potential relationship between curvature and types of relation.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 17399