Node- and edge-level performance of vision-based graph encoders

Determine whether vision-based approaches can match or exceed Graph Neural Networks on node-level and edge-level graph-learning tasks, where local structure matters more than in graph-level prediction.

Background

The survey reviews evidence that vision encoders processing rendered graph images can outperform GNNs on graph-level tasks requiring global structure understanding. However, the existing evidence is limited to graph-level prediction. Node-level and edge-level tasks place greater emphasis on local structure, and the survey identifies whether visual approaches can achieve performance comparable to or better than GNNs in these settings as unresolved.

References

However, current evidence remains limited to graph-level tasks. Whether vision-based approaches can match or exceed GNNs on node-level and edge-level tasks, where local structure matters more, remains an open question.

— When Vision Meets Graphs: A Survey on Graph Reasoning and Learning  (2609.03816 - Zhao et al., 3 Sep 2026) in Section 4, Vision for Graph Learning, subsection “Vision Encoders for Graph-Level Tasks”