{"id":"3933179b-a74c-4489-9f98-0a20686affc9","arxiv_id":"2412.11075","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AFECL learns graph representations by contrasting edges that share a node against all other edges, without data augmentation, and reports state-of-the-art results on low-label node classification and link prediction.","lead":"This paper presents a graph-learning method that trains on edges instead of nodes, treating edges that share an endpoint as similar and all other edges as different. It reports top accuracy for link prediction and for node classification when labeled examples are extremely scarce.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6) is undefined for non-neighbor negative pairs and does not implement the claimed 'other edges' contrast; the reported SOTA is not yet attributable to the stated loss.","rationale":"The central claim is that AFECL's edge-edge contrastive objective, Eq. (6), yields SOTA performance in low-label node classification and link prediction. Everything downstream, including Tables 2-5, the ablation, and the memory analysis, is evidence for that claim only if Eq. (6) is the objective actually optimized. That condition is the least secure part of the paper: hik is undefined for non-neighbor k, and the printed negative sums are not the 'other edges' described in the abstract. This is an internal definability and specification problem, not a disagreement with community consensus. The promised source code could settle it quickly, but the manuscript currently does not specify how the non-neighbor terms are computed. I also note secondary issues: Table 4 reports a standard deviation of 0.0 for AFECL on Cora despite five runs, and the duplicated anchor in Eq. (6) inflates the numerator; these are consistent with an under-specified or differently specified loss, but I do not treat them as independent objections. The reader's weakest assumption identified the same core issue, so I agree with the reader's judgment: conditional acceptance pending a precise statement and verification of the implemented loss. No adjustment to the reader's verdict is needed.","tokens_in":17772,"tokens_out":11000,"duration_ms":99888,"concrete_test":"Download the released code (github.com/YujunLi361/AFECL), locate the loss function, and reproduce Eq. (6) under two explicitly implemented variants: (A) denominator restricted to existing sampled edges, with negatives chosen from the sampled edge set; (B) denominator over all N node pairs by defining hik = W(hi∥hk) for non-edges. Record node classification accuracy on Cora/Citeseer at c=1 and link prediction AUC. If neither variant reproduces Tables 2 and 4, or if (A) and (B) differ significantly, the published equation is not the evaluated objective and the SOTA claim is unresolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (6) defines the proposed edge contrastive loss, but as written it is not computable. The denominator contains sums over all k ≠ i,j of exp(θ(hij,hik)/τ) and exp(θ(hij,hkj)/τ). The decomposition in the text immediately splits these into neighbor terms (vk ∈ Ni, vk ∈ Nj) and non-neighbor terms (vk ∉ Ni, vk ∉ Nj). However, hik and hkj are defined in Sec. 3.2 only for actual edges: the edge feature matrix is E ∈ R^{M×D}, and Eqs. (4)-(5) construct hij = W(hi∥hj) for (vi,vj) ∈ E. For k ∉ Ni, the pair (vi,vk) is not an edge, so hik has no definition. Thus either the loss was never evaluated as written, or the implementation silently defines node-pair embeddings for non-edges; the paper states neither. There is a second mismatch with the abstract and introduction: negatives are described as 'other edges,' but the negative terms in Eq. (6) are only hik and hkj, i.e., pairs sharing node i or j. Edges with no shared endpoint never enter the denominator, so the objective is not the global edge-edge contrast claimed. The anchor also appears three times in the numerator because j ∈ Ni and i ∈ Nj, while the denominator excludes it from the k-sums, so the normalization is internally inconsistent. Both issues are load-bearing because Eq. (6) is the entire pretext task; if the implemented loss differs from the printed one, the SOTA tables in Sec. 4 support a different method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes AFECL, an augmentation-free edge-level graph contrastive learning method. Node embeddings are learned by a multi-head GAT; an edge embedding is defined by concatenating the two endpoint node embeddings and applying a linear map or identity. The pretext task is an edge contrastive loss (Eq. 6) in which the anchor edge is contrasted with edges sharing one of its endpoints, with the paper claiming that edges sharing a node are positives and other edges are negatives. The authors report state-of-the-art results on semi-supervised node classification with 1-20 labels per class across eight datasets, link prediction on three datasets, and scalability experiments on Penn94 and ogbn-arxiv.","tokens_in":18147,"tokens_out":6998,"duration_ms":58844,"significance":"An augmentation-free, edge-level contrastive objective is a potentially valuable alternative to node-level GCL: it avoids handcrafted graph augmentations, exploits the comparatively rich edge set, and the concatenation-based edge embedding is computationally cheap. The paper provides code, extensive comparisons with 12 baselines, ablations, memory measurements, and hyperparameter sensitivity analyses. However, the central loss equation is under-specified: as written it references undefined embeddings and triple-counts the anchor, and the claimed global 'other edges' contrast is not what the equation implements. The empirical claims therefore currently support a method that is not fully described, and the contribution's validity depends on how these specification issues are resolved.","major_comments":[{"comment":"The denominator of Eq. (6) contains sums over all k not equal to i or j of exp(theta(hij,hik)/tau) and exp(theta(hij,hkj)/tau). Section 3.2 defines edge embeddings only for actual edges: E ∈ R^{M×D} and hij = phi(hi,hj) for (vi,vj) ∈ E. For k not in Ni, the pair (vi,vk) is not an edge, so hik is undefined. The paper therefore needs to state explicitly how the denominator is computed, for example by restricting the sums to existing edges or by defining node-pair embeddings for all pairs. If the implementation differs from the printed loss, the reported results must be re-evaluated under the implemented loss.","section":"Section 3.3, Eq. (6)"},{"comment":"Because j ∈ Ni and i ∈ Nj, the self-similarity term exp(theta(hij,hij)/tau) appears three times in the numerator, while the denominator excludes the anchor from both k-sums. The stated positive count |Ni|+|Nj|+1 therefore counts the anchor three times; the distinct number of positive pairs is |Ni|+|Nj|-1. The normalization in Eq. (6) is internally inconsistent and needs to be corrected, or the paper should justify the triple counting.","section":"Section 3.3, Eq. (6) and positive-pair count"},{"comment":"The paper repeatedly states that edges that do not share a node are negative pairs. However, the denominator of Eq. (6) only contains terms hik and hkj, meaning the contrast is limited to pairs of edges that share endpoint i or endpoint j; edges disjoint from {vi,vj} never enter the loss. If the intended contrast is global edge-edge, Eq. (6) must include all other edges; if the intended contrast is local, the abstract, introduction, and method description should be revised accordingly. This distinction affects the interpretation of the empirical results.","section":"Abstract, Section 1, and Section 3.3"},{"comment":"The complexity analysis states D'=KF' and derives O(N^2 KF') overall, but Eq. (5) defines W ∈ R^{D'×2KF'} and Table 8 reports values consistent with D'=2KF' (e.g., Cora with K=4 and F'=32 gives D'=256). The complexity claim therefore does not follow from the stated dimensions. In addition, Eq. (7) sums over all N^2 node pairs while the edge sampling procedure changes the number of nonzero terms to M', and the complexity analysis separately uses M' for large graphs; the relationship among Eq. (7), Algorithm 1, and the reported complexity needs to be reconciled.","section":"Section 3.3, Computational Complexity, and Table 8"}],"minor_comments":[{"comment":"The phrase 'optimizing the the edge contrastive loss' contains a duplicated article and should be corrected.","section":"Section 3.1"},{"comment":"The text 'edge–evel contrastive method' should read 'edge-level contrastive method'.","section":"Section 4.4"},{"comment":"The dataset name 'Chamelon' should be 'Chameleon', and 'Wikipeida' should be 'Wikipedia'.","section":"Appendix A.1"},{"comment":"The sentence 'the original graph faded into the encoder' should likely read 'the original graph fed into the encoder'.","section":"Section 2.2"},{"comment":"The sentence 'The results of homophilic and heterophilic graphs node classification accuracy are summarized in Table 2 and From the tables' is grammatically incomplete and should be rephrased.","section":"Section 4.4"},{"comment":"The description of the 'w/o ECL' variant is unclear; the sentence 'define edges corresponding to nodes connected by only one node as negative pairs' should be replaced with a precise statement of which pairs are positive and which are negative.","section":"Section 4.5"},{"comment":"The dataset descriptions state that Amazon-Photo nodes are 'documents' and Coauthor-CS nodes are 'documents'; these should be 'products' and 'authors', respectively.","section":"Section 4.1 and Appendix A.1"},{"comment":"The claim 'for the first time we study the edge-level pairs for contrast' should be tempered or supported with an explicit discussion of prior edge-level contrastive methods, since the phrase 'first' requires careful literature verification.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The GitHub repository was not examined during review. If the code is available, the authors should be asked to identify the exact loss implementation and to confirm whether it matches Eq. (6) as printed or a modified version. The resolution of this point is decisive for the reported results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: AFECL is the first augmentation-free edge-edge contrastive objective I know of, where edges sharing a node are positives. The edge embedding as a concatenation of node embeddings is cheap and sensible, and the idea of using edge-level topology for contrast in low-label regimes is worth taking seriously. The paper also does solid legwork on the experimental side: eight datasets, ablations, hyperparameter analysis, memory costs, and results on ogbn-arxiv. If the reported numbers hold up, the method is useful for scarce-label node classification and link prediction.\n\nNow the soft spots, and one is load-bearing. Equation (6), the entire pretext loss, is not well-defined as written. The denominator sums h_ik over all k not equal to i or j, but h_ik is defined only for actual edges. For non-neighbor k, that embedding does not exist. The decomposition in the text acknowledges non-neighbor terms explicitly, which means the authors must have implemented something else—probably restricting negatives to sampled edges or defining node-pair embeddings for non-edges. Neither is stated. On top of that, the anchor h_ij appears multiple times in the numerator because j is in N_i and i is in N_j, so the claimed positive count |N_i|+|N_j|+1 is wrong. And the objective is not the global \"other edges\" contrast promised in the abstract: only edges sharing the anchor's endpoints appear as negatives. These are not cosmetic typos; they change what is actually being optimized. Until the paper states the implemented loss, the SOTA tables support a different method.\n\nSmaller issues: Table 4 reports Cora link prediction as 96.5±0.0 over five runs, which is implausible and needs an explanation. The complexity analysis says D'=KF' while Eq. (5) and Table 8 use D'=2KF'. And the paper claims to avoid using validation labels for model selection, but the GAT hyperparameters must be chosen somehow; the procedure needs to be explicit.\n\nThat said, the core idea is sound and the flaws are fixable. The paper deserves a serious referee, but the referee should be instructed to verify what loss was actually implemented, ideally by checking the code. If the authors can provide a corrected equation and a clear implementation description, this could be a useful contribution. For now, I would not cite it as-is, but I would bring it to a reading group precisely because the gap between the written loss and the reported results is instructive.","headline":"Novel edge-level contrastive idea with an uncomputable loss equation as written; the strong empirical results cannot yet be attributed to the stated method.","tokens_in":18644,"tokens_out":2727,"would_cite":false,"duration_ms":25292,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that graph contrastive learning can be done at the level of edges, without augmentation, and that this edge-level contrast achieves state-of-the-art link prediction and extremely low-label node classification.","keywords":["graph contrastive learning","self-supervised learning","edge-level contrast","augmentation-free","link prediction","semi-supervised node classification","graph neural networks","low-label learning"],"falsifier":"Inspect the released code to check whether the denominator of Eq. (6) actually sums over all node pairs or only over sampled observed edges; if it is the latter, the implemented loss differs from the written one and the reported runtime and accuracy claims would need to be re-derived under the true loss. A separate check is to replace the shared-endpoint positive rule with random edge pairing on a heterophilic dataset and measure whether accuracy drops, which would reveal whether the topological contrast is the source of the reported gain.","tokens_in":17580,"feed_emoji":"🔗","tokens_out":8616,"duration_ms":69807,"temperature":0.7,"pith_summary":"This paper sets out to show that graph contrastive learning can be done without data augmentation and without moving to node-level or graph-level contrasts: the learning signal can live entirely on edges. The proposed model, AFECL, constructs an edge embedding by concatenating the embeddings of the edge's two endpoint nodes, then trains with a contrastive loss that treats edges sharing a node as positive pairs and all other edges as negative pairs. The claimed payoff is practical: a single-view, augmentation-free objective that produces strong representations for link prediction and for semi-supervised node classification when only one to four labeled nodes per class are available, and that uses less memory than comparable augmentation-free methods. A sympathetic reader would take the contribution to be the demonstration that edge-edge contrast, not node or view contrast, is a sufficient self-supervised objective for graphs.","feed_headline":"Edge contrasts beat graph augmentations when labels are scarce","feed_subtitle":"No augmentation, just edges: AFECL wins low-label node classification and link prediction.","key_machinery":"The central object is the edge embedding $h_{ij} = \\varphi(h_i, h_j)$, obtained by concatenating the endpoint node embeddings $h_i$ and $h_j$ through an identity or learnable map $W$, together with the edge contrastive loss $\\ell(h_{ij})$ in Eq. (6). This loss is an InfoNCE-style ratio: the numerator contains the anchor edge and all edges incident to either endpoint (positives), and the denominator contains those positives plus non-incident edges (negatives), so topology enters the objective directly through the choice of pairs. Edge sampling via $A' = A \\circ R$, with $R_{ij} \\sim \\mathrm{Bernoulli}(p_s)$ on existing edges, is the mechanism that keeps the $O(N^2 K F')$ overall complexity manageable on large graphs.","core_discovery":"On its own terms, the paper's discovery is that edges can be the contrastive instances in graph contrastive learning, and that the original unmodified graph is enough to define them: no corruptions, augmentations, or second view are required. The model defines $h_{ij} = \\varphi(h_i, h_j)$, the embedding of edge $(v_i, v_j)$ as the (possibly learned) concatenation of its endpoint node embeddings, and optimizes a loss in which an anchor edge $h_{ij}$ is pulled toward itself and toward edges incident to either of its endpoints, while being pushed away from non-incident edges. Across eight datasets, the paper reports that this objective reaches state-of-the-art or competitive accuracy relative to twelve baselines, including two supervised GNNs, in low-label node classification, achieves the best AUC on link prediction for Cora, Citeseer, and Coauthor-CS, and scales to large graphs by sampling edges rather than constructing edge neighborhoods.","pith_inferences":["Extension: the paper's positive-pair rule is motivated by topology alone, but the paper does not analyze when shared-node edges are actually semantically similar; on heterophilic graphs the reported gains suggest the objective may be capturing structural or positional information beyond label homophily, which would be worth isolating with synthetic graphs.","Extension: a direct testable consequence is that the denominator's treatment of non-neighbor edges determines both memory and behavior; checking the released implementation against Eq. (6) would settle whether the written all-pairs loss is what was evaluated.","Extension: the same edge-contrastive loss could be probed as a pretraining objective for heterogeneous or temporal graphs, where edge types carry meaning; the paper only tests the fixed homogeneous-graph setting."],"forward_implications":["No augmentation tuning: because AFECL uses the original graph as its only view, users do not need to select dataset-specific graph transformations such as edge dropping or attribute masking.","Low-label gains: with one to four labeled nodes per class, AFECL reports the best or second-best accuracy on Cora, Citeseer, PubMed, Coauthor-CS, and Amazon-Photo among the compared methods.","Direct link prediction strength: since edge embeddings are endpoint concatenations, they can be scored for missing links; AFECL reports higher AUC than the compared methods on Cora, Citeseer, and Coauthor-CS.","Comparable cost to node-node GCL: the paper's complexity analysis puts AFECL at $O((NFF' + N^2F')K)$, the same order as representative node-node methods, and its measured memory use is lower than SPGCL and GraphACL on citation graphs.","Scalability to large graphs: edge sampling plus the single-view design lets AFECL run on Penn94 and ogbn-arxiv, where it matches or exceeds the reported baselines."],"supporting_citations":[{"why":"Existing augmentation-free node-level contrastive method that AFECL extends to edge-level and compares against.","marker":"(Xiao et al. 2024)"},{"why":"Source of the neighbor-contrastive positive-pair idea and of the extremely-scarce-label (1-4 per class) evaluation protocol.","marker":"(Shen et al. 2023)"},{"why":"GRACE baseline; representative node-node augmentation-based GCL method used for comparison and for complexity benchmarking.","marker":"(Zhu et al. 2020)"},{"why":"DGI baseline; augmentation-free node-graph contrastive method that AFECL's single-view design contrasts with.","marker":"(Veli\\v{c}kovi\\'{c} et al. 2019)"},{"why":"The GAT encoder whose attention coefficients generate the node embeddings from which edge embeddings are concatenated.","marker":"(Veli\\v{c}kovi\\'{c} et al. 2017)"},{"why":"Supplies the transductive link prediction protocol: train encoder without labels, freeze it, then train a decoder.","marker":"(Shiao et al. 2023)"},{"why":"Provides the HLCL baseline and Penn94 experimental setting used to test AFECL on a large heterophilic graph.","marker":"(Yang and Mirzasoleiman 2024)"}],"fun_headline_variants":["Edge contrasts without augmentation top scarce-label tasks","AFECL makes edges the contrastive pairs, no corruptions","Scarce labels? Edge-level contrast beats GCL baselines","No augmentation, just edge contrasts: new SOTA for graphs","Link prediction and node classification via edge contrasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the original graph's topology alone defines a useful notion of contrastive similarity—edges that share an endpoint should be pulled together and every other edge pair pushed apart—so that the objective remains beneficial across both homophilic and heterophilic graphs without any augmentation.","fun_headline_variants_meta":{"raw":{"variants":["Edge contrasts without augmentation top scarce-label tasks","AFECL makes edges the contrastive pairs, no corruptions","Scarce labels? Edge-level contrast beats GCL baselines","No augmentation, just edge contrasts: new SOTA for graphs","Link prediction and node classification via edge contrasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1496,"prompt_tokens":957,"completion_tokens":539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":573,"tokens_out":539,"duration_ms":4629,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:19:36.237744+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code to check whether the denominator of Eq. (6) actually sums over all node pairs or only over sampled observed edges; if it is the latter, the implemented loss differs from the written one and the reported runtime and accuracy claims would need to be re-derived under the true loss. A separate check is to replace the shared-endpoint positive rule with random edge pairing on a heterophilic dataset and measure whether accuracy drops, which would reveal whether the topological contrast is the source of the reported gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Existing augmentation-free node-level contrastive method that AFECL extends to edge-level and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the neighbor-contrastive positive-pair idea and of the extremely-scarce-label (1-4 per class) evaluation protocol."},{"cited_title":"L.; Liò, P.; Bengio, Y.; and Hjelm, R","cited_arxiv_id":null,"evidence_quote":"DGI baseline; augmentation-free node-graph contrastive method that AFECL's single-view design contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the transductive link prediction protocol: train encoder without labels, freeze it, then train a decoder."}],"review_version":1}