Research map / Graph learning / GNN architectures

Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction

ICML 2025 · Graph learning › GNN architectures · approach: Graph rewiring and scalable message passing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Renders k-hop subgraphs around links as images, extracts visual structural features with a pretrained ResNet50, and fuses them with MPNNs (GCN, NCNC) for link prediction, with an efficient node-centric variant E-GVN.

Problem

MPNN-based link prediction relies on message passing and heuristic structural features, which struggle to distinguish links involving isomorphic nodes and to count substructures; the potential of visual perception of graph structure has not been exploited.

Approach

Render each link's local k-hop subgraph as an image with a graph visualizer, encode it with a vision encoder into visual structural features (VSFs), and fuse them with MPNN node representations via cross-attention; E-GVN instead visualizes per-node subgraphs once, adapts the frozen VSFs with a trainable adapter, and fuses them into node attributes before message passing.

Key results

• GVN_GCN raises GCN HR@100 on Cora from 66.79 to 81.13 and on Pubmed from 53.02 to 73.17.
• E-GVN_NCNC reaches 91.47 HR@100 on Cora and 84.02 on Pubmed, versus NCNC's 89.65 and 81.29, establishing new SOTA on the reported benchmarks.
• E-GVN_NCNC improves ogbl-ppa HR@100 from 61.42 to 63.45 and ogbl-ddi HR@20 from 84.11 to 87.31 over NCNC.

Measured effects reported in the paper

Surprising findings

• Pure-VSF features (ResNet50 of subgraph images) outperform 2-D coordinates and Laplacian positional encodings used as node features, showing the visual encoding itself matters.
• Removing node labels from visualized subgraphs performs best; adding unique global IDs does not help, and a 'no-label' setting beats re-labeling.

Limitations

• GVN requires per-link visualization and is too time-consuming for large-scale graphs (over 12 hours per epoch on some datasets, OOM for some baselines).
• VSFs are not permutation-equivariant; the authors justify their effectiveness heuristically rather than theoretically.
• Many GVN results on large OGB datasets are missing because of computational cost.

Experimental setup

Datasets: Cora Citeseer Pubmed ogbl-collab ogbl-ppa ogbl-citation2 ogbl-ddi

Benchmarks: Cora Citeseer Pubmed ogbl-collab ogbl-ppa ogbl-citation2 ogbl-ddi

Metrics: HR@100 HR@50 HR@20 MRR hit@1 hit@3 hit@10 hit@50

Baselines: CN AA RA GraphSAGE GCN SEAL NBFNet Neo-GNN BUDDY NCNC

Builds on: NCNC SEAL GCN Graphviz

Closest papers by meaning

• Structural Information Enhanced Graph Representation for Link Prediction (AAAI 2024 · similarity 0.89)
• The Underappreciated Power of Vision Models for Graph Structural Understanding (NeurIPS 2025 · similarity 0.89)
• Vision Graph Prompting via Semantic Low-Rank Decomposition (ICML 2025 · similarity 0.87)
• Pure Message Passing Can Estimate Common Neighbor for Link Prediction (NeurIPS 2024 · similarity 0.87)
• A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.86)
• On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.86)