{"id":"d99f40f0-3ab8-4e92-877b-18532aff45c1","arxiv_id":"2505.08266","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Visual features extracted from rendered subgraph images consistently improve message-passing link predictors on seven benchmarks, and an efficient node-centered variant scales to large graphs.","lead":"This paper proposes rendering the neighborhood around a candidate link as an image and feeding that image through a standard vision model, then combining the visual features with a graph neural network for link prediction. Across seven graph benchmarks, the vision-enhanced models consistently beat their base models and reach new best results among the methods compared.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GVN/E-GVN may leak the target link through the layout: the queried edge is masked visually but not removed from the force-directed layout, so node coordinates encode the label.","rationale":"The reader's verdict ACCEPT rested on the assumption that a force-directed rendering preserves enough topology for useful structural features. I find a more specific and more damaging issue: target-label leakage through the layout. The paper's own text and algorithm listings do not establish that the queried edge is removed from the graph before visualization; 'masked' in Section 4.1 is ambiguous and likely means hidden in the rendered image, not absent from the layout computation. Force-directed layouts are deterministic functions of the full edge set, so the presence or absence of the queried edge changes node coordinates, which a ResNet can read. E-GVN lacks even that masking language. This is an internal consistency concern, not a disagreement with consensus: the method's own descriptions admit a shortcut that would trivially encode the label. If the shortcut is present, the central claim is unsupported by the current experiments; if the code shows the edge is deleted before layout, the concern vanishes and the empirical case is reasonably strong. Hence I recommend CONDITIONAL: require the authors to verify and, if needed, re-run with target-edge-free layouts. The reader's assumed weakest point (layout distorts topology) is related but not the same; the leakage is the load-bearing risk.","tokens_in":22150,"tokens_out":11934,"duration_ms":125822,"concrete_test":"Run the official code and inspect the subgraph construction: does the queried edge (u,v) remain in the graph passed to Graphviz? Then perform two controlled experiments on Cora and Citeseer with GVNNCNC and E-GVNNCNC: (1) re-render all images with the target edge removed from the graph before layout (for E-GVN, remove the queried edge per link or use an edge-free graph at inference), and compare Hits@100; (2) compute the correlation between the Euclidean distance of the two colored endpoints in the rendered image and the binary label on a held-out split. If the distance-label correlation is high or performance drops sharply when the edge is removed from the layout, the visual gains are confounded by leakage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that visual structural features (VSFs) extracted from rendered subgraph images improve MPNN link prediction. The load-bearing condition is that the rendered image does not contain the answer to the link prediction question. Section 4.1 states only that 'the link itself is masked for prediction,' which is a rendering operation; Algorithm 1 extracts the k-hop subgraph S^k_uv from G and passes it to the visualizer without deleting (u,v). Graphviz's sfdp is a force-directed layout: if the edge (u,v) is present in the graph object, it pulls the two colored endpoint nodes together, so the ResNet can infer the label from their spatial proximity even when the edge is not drawn. For E-GVN (Algorithm 3), the situation is worse: subgraphs are node-centered and no masking is described, so a positive training edge (u,v) is actually rendered as a drawn edge in the image of v; and on ogbl-collab, validation edges are explicitly added to G at test time (Appendix D), making the shortcut available during validation. This would explain the large gains on ogbl-collab (38.86% relative for E-GVNGCN) and the near-perfect substructure counting in Table 1, where the encoder can simply read edge presence. All ablations (styles, visualizers, encoders, integration) share this potential leak, so they do not control for it. If the target edge is not removed before layout, the reported improvements reflect label leakage, not generalizable visual structural awareness.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Graph Vision Network (GVN) and its efficient variant E-GVN, which render k-hop subgraphs around query links (or around nodes) as images via Graphviz/sfdp, encode them with a pretrained ResNet50 to obtain Visual Structural Features (VSFs), and fuse these into MPNNs (GCN, NCNC) for link prediction. The authors report consistent gains across seven datasets, including OGB large-scale graphs, and claim new SOTA results. The paper includes ablations on visualization scope, style, encoder, integration strategy, and a substructure-counting analysis.","tokens_in":22428,"tokens_out":6180,"duration_ms":61196,"significance":"If the reported gains reflect genuine visual-structure awareness rather than leakage of the target-edge label into the rendered image, the work would open a new and plausibly orthogonal direction for link prediction, with practical value from the efficient node-centered variant. The manuscript is generally well organized, releases code, and provides detailed ablations. However, as argued in the major comments, the current experimental protocol does not rule out a direct visual shortcut, and the substructure-counting evidence is consistent with the encoder reading edge presence. The contribution is therefore significant in potential but not yet established.","major_comments":[{"comment":"The central claim that VSFs encode link-relevant structure is threatened by a label-leakage mechanism. In Algorithm 1 (GVN), Step 1 extracts the k-hop subgraph S^k_uv from G and passes it to the visualizer without deleting the query edge (u,v); Section 4.1 only states that the link is 'masked for prediction,' which is a drawing operation. Since the default layout is Graphviz's force-directed sfdp, the presence of (u,v) in the graph object used for layout pulls the two highlighted endpoint nodes together, so the rendered image encodes the label through spatial proximity even when the edge stroke is not drawn. For E-GVN (Algorithm 3), the node-centered subgraph S^k_v is visualized with no masking at all, so a positive training edge (u,v) is rendered as a drawn edge inside the image of v, whereas a negative pair has no such edge; the model can therefore read the label directly from image content. Appendix D further states that on ogbl-collab validation edges are added to G at test time, making the shortcut available in validation. This leakage would explain the large gains in Table 3 (e.g., 38.86% relative improvement for E-GVNGCN on ogbl-collab) and the near-perfect substructure counts in Table 1. The authors must either remove the target edge from the subgraph before layout (for both GVN and E-GVN) or otherwise demonstrate that the layout and rendering do not convey the target edge's existence; without such a control, the empirical core of the paper does not support RQ2.","section":"Section 4.1, Algorithms 1 and 3, Appendix D"},{"comment":"The substructure-counting experiment is presented as evidence that VSFs confer fine-grained substructure awareness, but the near-zero normalized MSE values (e.g., 6.76E-9 for VSF+GCN) indicate that the vision encoder essentially reads the edge set directly from the rendered graph. This is not evidence of a generally useful visual structural feature; it is exactly what one would expect if the image contains the adjacency information. To support Remark 4.2, the substructure-counting task should be run with a protocol where the image is the only input and the target substructure is not trivially readable as drawn edges, or the experiment should be reinterpreted as a demonstration of edge-presence reading rather than structural generalization.","section":"Table 1 and Section 4.2"},{"comment":"The claim that GVN/E-GVN 'achieve new SOTA results' is stronger than the evidence supports. Baseline numbers are taken directly from Wang et al. (2024) rather than re-run under the authors' protocol; as Appendix D notes, the treatment of validation edges as message-passing paths differs across datasets and baselines. Even setting aside the leakage issue, the comparison would benefit from first-hand baseline runs and standard significance tests; at least one overlap of the reported mean and standard deviation (e.g., E-GVNNCNC vs NCNC hit@100 on Cora, 91.47±0.36 vs 89.65±1.36) is not formally assessed.","section":"Section 5.1 and Table 3"}],"minor_comments":[{"comment":"The message-passing update equation has unbalanced parentheses: the expression should be corrected to clearly delimit the aggregation set and the update inputs.","section":"Section 3, Eq. (1)"},{"comment":"The header of Table 7 contains a typo: 'Igraj' should be 'Igraph'.","section":"Table 7"},{"comment":"The terms 'Adaptor' and 'Adapter' are used interchangeably; please unify the spelling throughout.","section":"Section 4.4, Algorithms 2 and 3"},{"comment":"The header of Table 12 lists 'Collab,PPA,Collab,DDI' with 'Collab' appearing twice; it should be 'Collab, PPA, DDI, Citation2'.","section":"Appendix C, Table 12"},{"comment":"The comparison with 2-dimensional coordinates uses raw coordinates fed into a GCN, not through the vision encoder; clarify that this comparison conflates encoder capacity with information content, and consider a control that passes the same image without the queried-edge layout cue.","section":"Section 5.5, Table 11"}],"recommendation":"major_revision","confidential_remarks":"The leakage issue raised in Major Comment 1 is decisive: if the target edge is not removed before layout, the reported improvements can be explained by a shortcut. I recommend requesting a revision that reruns all main experiments with the query edge (or, for E-GVN, the queried neighbor relation) excluded from the graph object passed to the visualizer, and that reports whether the gains persist. If the authors cannot provide such results, a reject recommendation would be appropriate. The writing and organization are otherwise fine; the concern is empirical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is novel, but the main experiments are likely compromised by label leakage. In Algorithm 1, the k-hop subgraph S^k_uv is extracted from G without deleting the queried edge (u,v). Section 4.1 says 'the link itself is masked for prediction,' but that only affects rendering; the sfdp layout still sees the edge and pulls the two brown endpoints together. A ResNet can exploit that spatial proximity. For E-GVN (Algorithm 3), it is worse: node-centered images of v will actually draw the edge to u for positive training links, so the VSF for v contains the answer. And on ogbl-collab, validation edges are added back to G at test time (Appendix D), so the shortcut is available for model selection. This explains the large relative gains on ogbl-collab (38.86% for E-GVNGCN) and would make the ablations uninformative because every variant shares the leak.\n\nWhat's genuinely new: the framework is the first I know to render k-hop subgraphs and feed a pretrained vision encoder into MPNN link prediction. The paper is careful about style consistency, encoder choice, scope, and integration strategies, and the large-scale variant is a sensible way to cut cost from O(l) to O(n). If the leak is fixed, this could be a useful direction.\n\nThe substructure counting in Table 1 is not the same leak—the image is the whole graph, not a queried edge—but the near-perfect errors (1e-8) are surprising and deserve a closer look; a vision encoder counting triangles that accurately seems too good without some other confound.\n\nBottom line: the central empirical claim is not supported as written. The fix is straightforward—remove the target edge from the subgraph before layout, and for E-GVN mask or remove all edges incident to the query node pair appropriately—and rerun with a control that the image alone cannot predict the label. I would send this to review because the idea is worth engaging with, but I would not accept without the leakage being addressed and new numbers.\n\nRecommendation: major revision with a clean-room reproduction of the pipeline.","headline":"The paper's gains likely come from the queried edge remaining in the visual layout; the leak must be fixed before the claim is credible.","tokens_in":22985,"tokens_out":4129,"would_cite":false,"duration_ms":42346,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Vision features extracted from rendered subgraph images improve MPNN link prediction across all seven benchmarks and set new state-of-the-art results.","keywords":["link prediction","message passing neural networks","visual structural features","subgraph visualization","vision encoder","graph neural networks","multimodal graph learning","structural features"],"falsifier":"Run GVN with all rendering settings fixed except the layout: replace the force-directed sfdp placement with random node positions on the same canvas. If link-prediction accuracy stays at the same level, the encoder is reading pixel artifacts rather than structure and the gains would not generalize across renderers; if accuracy collapses, layout fidelity is load-bearing. A second check: replace the pretrained ResNet50 with a randomly initialized encoder of the same architecture; if the gains survive, natural-image pretraining is not the source of the VSF signal.","tokens_in":21946,"feed_emoji":"👁️","tokens_out":9004,"duration_ms":83237,"temperature":0.7,"pith_summary":"The paper's central claim is that a message-passing link predictor can see better when it literally looks: render the k-hop subgraph around each candidate link as a consistent-style image, read that image with a pretrained vision encoder, and fuse the resulting visual structural features into the MPNN. On this basis the authors propose the Graph Vision Network (GVN) and an efficient variant (E-GVN), and report that the visual channel improves link prediction across all seven datasets tested, including four large-scale OGB benchmarks, while remaining a plug-in compatible with existing state-of-the-art models. If the claim holds, vision becomes an orthogonal source of structural signal, capable of distinguishing links that message passing and hand-designed structural features cannot, and a broad new direction for graph learning opens up.","feed_headline":"Feeding graph images to a vision encoder boosts link prediction","feed_subtitle":"Rendered k-hop subgraphs give MPNNs visual structural cues that message passing and hand-built features miss.","key_machinery":"The load-bearing object is the Visual Structural Feature (VSF): the output vector of a vision encoder (default ResNet50) applied to a fixed-style image of the k-hop subgraph enclosing the candidate link (GVN) or centered on a node (E-GVN), with the queried endpoints colored, labels removed, and the link itself masked. The rendering pipeline, subgraph extraction, Graphviz sfdp layout, and consistent style, converts topology into pixels in a way the encoder can read; the visual perception scope $k \\le 3$ is deliberately decoupled from message-passing depth, so the model sees a crisp local picture while the MPNN reaches wider. The framework's other moving parts are the three integration strategies (attention-based, concatenated, and weighted) that inject VSFs without touching the message-passing loop, and, in E-GVN, the node-centered visualization that cuts rendering cost from $\\mathcal{O}(l)$ per-link to $\\mathcal{O}(n)$ per-node plus a frozen-encoder/adaptor design that keeps memory flat.","core_discovery":"GVN and E-GVN establish, empirically, that visual structural features (VSFs), vectors produced by a ResNet50 encoder applied to a force-directed rendering of a k-hop subgraph, carry genuine link-relevant information that MPNN representations and classical structural features do not fully cover. The paper demonstrates this three ways: link-centered images of subgraphs with isomorphic endpoints are visibly different, so VSFs discriminate links that the 1-WL-limited MPNN equates; VSFs let plain GCN and SAGE count triangles and 3-stars with near-zero normalized error in synthetic substructure-counting tests; and VSFs reproduce the values of six standard structural features (CN, RA, AA, SPD, DRNL, DE), with the reproduced mix shifting by dataset density after finetuning. Integrated into MPNNs by cross-attention, concatenation, or weighted prediction, VSFs improve GCN by 21 to 38 percent relative on the citation networks and push the state-of-the-art model NCNC higher on every one of the seven datasets, which the paper reads as evidence that vision awareness is orthogonal to common-neighbor and path-based features. E-GVN moves rendering from per-link to per-node, freezes the encoder, and appends a trainable adapter, which brings the same benefits to large-scale graphs at near-base-model cost.","pith_inferences":["The adaptivity result (Remark 4.4) implies a transfer experiment the paper does not run: an encoder tuned on a dense graph should, via the adapter, re-weight toward path-based features when moved to a sparse graph, which would turn VSFs into a reusable cross-dataset structural-feature layer.","The same render-and-encode pipeline is a natural fit for node classification, graph classification, and heterogeneous or temporal graphs, where the local visual pattern carries information distinct from node attributes; the paper's own impact statement flags these extensions.","Since VSFs are permutation-sensitive, rendering effectively injects a data-augmentation family (layout seeds, rotations, flips, colors) into a permutation-equivariant model; a direct follow-up is to measure which of these visual augmentations regularize learning best.","A cheap control suggested by the partial-adaptivity result: replace the pretrained ResNet50 with a random-weight encoder and keep the adapter trainable. If accuracy holds, pretrained visual knowledge is not the source of the gain; if it collapses, natural-image pretraining is load-bearing."],"forward_implications":["Vision enhancement lifts plain MPNNs: GVN and E-GVN over GCN give relative HR@100 gains of about 21%, 25%, and 38% on Cora, Citeseer, and PubMed respectively.","The gains stack on top of the strongest common-neighbor-based model: GVN and E-GVN with NCNC beat NCNC on all seven datasets, both on the headline metric and on most of the broader hit@k and MRR metrics.","VSFs supply MPNNs with substructure-counting ability: adding VSFs drops normalized counting error for triangles and 3-stars from order-1 values to below $10^{-5}$ in the synthetic tests.","Because the visual scope is decoupled from message-passing depth, the model gets both a refined local structural readout and a wider message-passing reach, so the design does not sacrifice depth for vision.","E-GVN's node-centered, frozen-encoder design makes the approach usable on large-scale graphs: relative gains of 38.86% (HR@50) on ogbl-collab, 72.20% (HR@100) on ogbl-ppa, 1.60% (MRR) on ogbl-citation2, and 62.42% (HR@20) on ogbl-ddi over plain GCN."],"supporting_citations":[{"why":"Supplies the ResNet50 vision encoder that extracts the visual structural features in both GVN and E-GVN.","marker":"He et al., 2016"},{"why":"Supplies Graphviz, the default graph visualizer whose sfdp layout renders each k-hop subgraph into the input image.","marker":"Gansner & North, 2000"},{"why":"NCNC is the state-of-the-art SF-enhanced MPNN that GVN/E-GVN wrap and the strongest baseline their results are measured against.","marker":"Wang et al., 2024"},{"why":"SEAL contributes the k-hop enclosing-subgraph extraction and DRNL path-based structural features that frame the subgraph and SF design.","marker":"Zhang & Chen, 2018"},{"why":"BUDDY is the subgraph-sketching state-of-the-art baseline whose common-neighbor features the visual features are shown to complement.","marker":"Chamberlain et al., 2023"},{"why":"Supplies the four large-scale OGB benchmarks (collab, ppa, citation2, ddi) used to test scalability.","marker":"Hu et al., 2020"},{"why":"Provides the decoupling-of-depth-and-scope rationale that justifies fixing the visual perception scope k.","marker":"Zeng et al., 2021"},{"why":"Supplies the substructure-counting experimental setup used to show VSFs give MPNNs fine-grained motif awareness.","marker":"Chen et al., 2020"},{"why":"GCN is the base message-passing model in the GVNGCN and E-GVNGCN configurations.","marker":"Kipf & Welling, 2017"}],"fun_headline_variants":["Visual subgraph images boost link prediction in MPNNs","Graph vision improves link prediction to new SOTA","ResNet on graph images sharpens link prediction","Vision-aware message passing sees links better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"A force-directed drawing of a small subgraph, read by an image encoder pretrained on natural photographs, preserves enough of the topology that matters for links (shared neighbors, distances, motifs) that the encoded features genuinely help predict links rather than just re-encoding drawing artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Visual subgraph images boost link prediction in MPNNs","Graph vision improves link prediction to new SOTA","ResNet on graph images sharpens link prediction","Vision-aware message passing sees links better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1562,"prompt_tokens":943,"completion_tokens":619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":560}},"tokens_in":559,"tokens_out":619,"duration_ms":6309,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:58:58.479536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GVN with all rendering settings fixed except the layout: replace the force-directed sfdp placement with random node positions on the same canvas. If link-prediction accuracy stays at the same level, the encoder is reading pixel artifacts rather than structure and the gains would not generalize across renderers; if accuracy collapses, layout fidelity is load-bearing. A second check: replace the pretrained ResNet50 with a randomly initialized encoder of the same architecture; if the gains survive, natural-image pretraining is not the source of the VSF signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Graphviz, the default graph visualizer whose sfdp layout renders each k-hop subgraph into the input image."},{"cited_title":"Neural common neighbor with completion for link prediction","cited_arxiv_id":null,"evidence_quote":"NCNC is the state-of-the-art SF-enhanced MPNN that GVN/E-GVN wrap and the strongest baseline their results are measured against."},{"cited_title":"and Chen, Y","cited_arxiv_id":null,"evidence_quote":"SEAL contributes the k-hop enclosing-subgraph extraction and DRNL path-based structural features that frame the subgraph and SF design."},{"cited_title":"P., Shirobokov, S., Rossi, E., Frasca, F., Markovich, T., Hammerla, N","cited_arxiv_id":null,"evidence_quote":"BUDDY is the subgraph-sketching state-of-the-art baseline whose common-neighbor features the visual features are shown to complement."},{"cited_title":"Decoupling the depth and scope of graph neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the decoupling-of-depth-and-scope rationale that justifies fixing the visual perception scope k."},{"cited_title":"Can graph neural networks count substructures? Advances in Neural Information Processing Systems, 33: 0 10383--10395, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the substructure-counting experimental setup used to show VSFs give MPNNs fine-grained motif awareness."}],"review_version":1}