GRIPedge: Heterophily-aware graph learning via attentional feature-spectral neighbour propagation in dense graphs

ICLR 2026 Conference Submission21145 Authors

Keywords: graph learning, node classification, heterophilic dense graphs

Abstract:

Node classification in heterophilic graphs remains a challenging task, as connected
nodes often belong to different classes and exhibit heterogeneous features. The assumption
of homophily, which is typical in GNNs, encounters problems such as
oversmoothing and reduced separability and leads to low performance on dense
heterophilic benchmarks such as Squirrel and Chameleon. We therefore propose
a unified framework that improves feature representation, structural learning and
spectral aggregation. Our approach combines attention-based mechanisms to integrate
local and global neighborhood information, spectral modulation to capture
oscillatory node–edge patterns and edge augmentation inspired by structure
learning to refine graph connectivity. Extensive experiments demonstrate that our
model consistently offers robust and discriminative node embeddings and outperforms
state-of-the-art methods on the task of node classification in dense graphs.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 21145