Browse Teach Curate

K-Path Rooted Subgraph Expressivity for Graph-Level Graph Neural Networks

Raises the expressive power of message-passing Graph Neural Networks on graph-level tasks by rooting each node's representation in its k-path subgraph, enabling the model to count substructures such as paths and cycles that standard GNNs cannot distinguish. Paired with an adaptive graph contrastive-learning augmentation that removes unimportant edges to improve cross-domain generalization.

Rosie Text

2501.00773

This paper studies graph-level learning with Graph Neural Networks (GNNs) - models that predict a property or class for an entire graph, as in molecular property prediction and subgraph counting. It …