{"id":"549665f8-e895-44c3-acd8-62fcb8c279ec","arxiv_id":"2502.02302","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"EdgeGFL multiplies node messages by learned edge-type vectors to implement per-dimension feature preference in heterogeneous graph neural networks, reporting small gains over prior GNN baselines.","lead":"This paper proposes EdgeGFL, a graph neural network that uses learned edge-type embeddings to filter or amplify individual feature dimensions of messages passed between nodes. The authors report state-of-the-art node classification and clustering results on four heterogeneous graph benchmarks, but the method is under-specified and the experiments lack code and error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. (9)-(10) and (11)-(13) define conflicting message-passing schemes, and Eq. (13)'s attention score is never defined; the reported SOTA margins cannot be verified without resolving which EdgeGFL instantiation produced Table III.","rationale":"The reader correctly identified the undefined attention computation in Eq. (13) as a load-bearing under-specification. My stress-test goes further: the under-specification is accompanied by a direct contradiction between the message-passing definition in Eqs. (9)-(10) and the aggregator in Eqs. (11)-(13). The first uses per-dimension edge vectors r^l_ij and no trainable projection; the second uses scalar attention weights α^l_ij and a trainable projection W^l, with no role for r^l_ij. Algorithm 1 explicitly routes execution through Eqs. (9)-(10), while the ablation study and contribution text describe the edge-dot-product mechanism. This is not merely a missing formula; it means the evaluation in Table III could correspond to either model, or to a third model that is not described. The complexity analysis in Section IV-F, which claims O(n) per layer despite dot products over all edges, reinforces that the implementation details are not reliable. Because the core performance claim—outperforming SeHGNN by 1.51% average micro-F1—cannot be checked against any uniquely specified method, the paper is not acceptable in its current form. I agree with rejection, but I would ground it primarily in the internal inconsistency of the method definition rather than only in the missing α-hat formula.","tokens_in":24062,"tokens_out":3528,"duration_ms":35693,"concrete_test":"Obtain the authors' code or reimplement both possible instantiations from the text: (A) the literal Eqs. (8)-(10) model, with edge-gated messages summed without residual or attention, and (B) the Eqs. (11)-(13) model, with α-hat computed by the natural but unspecified edge-type-plus-node attention MLP and residual α. Train each on the four datasets with the stated 24/6/70 split, output L2 normalization, and the same loss as Eq. (16). If neither variant reproduces EdgeGFL's Table III micro-F1 values (DBLP 95.98, ACM 94.42, IMDB 69.36, Freebase 69.04) within one standard deviation, the SOTA claim rests on an unavailable instantiation; if one matches, the paper should state that variant explicitly and replace the contradictory Eqs. (9)-(10) description.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires a well-defined EdgeGFL instantiation, but the manuscript provides two inconsistent ones. Section IV-C defines message propagation as per-dimension edge gating, Eq. (9): M^l_ij = h^l_j · r^l_ij, followed in Eq. (10) by a plain sum aggregator with no residual, no activation, and no projection. Section IV-D then defines the implemented aggregator as Eq. (12): h^{l+1}_i = σ(Σ_{j∈N_i} α^l_ij W^l h^l_j + W^l_res h^l_i), a scalar-attention GCN with residual connections in which the edge vector r^l_ij never appears. Algorithm 1 instructs the reader to use Eqs. (9)-(10), while the contribution narrative and the ablation description (w/o FGL: 'no edge-to-node mapping or dot product') describe Eq. (9). These are different models. The attempted bridge, Eq. (13), applies residual connections to attention scores but never defines α-hat, saying only that it 'employs both edge type embeddings and node embeddings.' Consequently, the reported margins over SeHGNN in Table III—an average of 1.51% micro-F1, and the 5.38% Freebase gain—are attached to an algorithm that is not uniquely specified. Correctness risk is high not because the idea is implausible but because the two explicit formulations conflict and no unique model can be evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EdgeGFL, a heterogeneous graph neural network layer in which each edge type is represented by a learnable multidimensional vector that scales individual feature dimensions of propagated messages, and which combines this with residual connections and attention-based aggregation. The authors evaluate node classification (micro-F1/macro-F1), clustering (ARI/NMI and Precision/MAP), embedding visualization, and parameter sensitivity on DBLP, ACM, IMDB, and Freebase, and report state-of-the-art results over SeHGNN and other baselines. The central claimed contribution is that per-dimension edge gating improves node representations compared to scalar or attention-only weighting.","tokens_in":24414,"tokens_out":10303,"duration_ms":93032,"significance":"The multidimensional edge-gating idea is a plausible and potentially useful incremental contribution over scalar or type-level relation weighting in heterogeneous graph learning, and the experimental coverage (four benchmarks, multiple baselines, ablations, parameter sensitivity, and reported p-values) is a strength. However, the manuscript currently does not define the exact model that produced the tables: the propagation and aggregation equations describe conflicting instantiations, and the attention score used in the implemented aggregator is never specified. Until these definitions are supplied and the reported summary statistics are reconciled with the tables, the claimed state-of-the-art result cannot be credited. No code or data-release statement is provided, which further limits reproducibility.","major_comments":[{"comment":"The manuscript specifies at least two incompatible message-passing instantiations. Equation (9) defines the message as the product of the source node representation h^l_j and the edge preference vector r^l_ij, followed in Eq. (10) by a plain sum aggregator; Algorithm 1 instructs the implementation to use Eqs. (9)-(10). In contrast, Eqs. (11)-(12) define an attention-weighted aggregation with a residual connection in which the edge vector r^l_ij does not explicitly appear, and Eq. (13) adds a residual connection to attention scores. As written, the reader cannot determine which aggregation produced Table III, so the central claim that EdgeGFL outperforms SeHGNN by 1.51% and 0.64% is not attached to a uniquely specified algorithm. The revision must either unify these formulations or explicitly state which variant generated each reported result.","section":"IV-C, IV-D, Algorithm 1"},{"comment":"The attention coefficient alpha-hat in Eq. (13) is never defined. The sentence following the equation only says that it employs both edge type embeddings and node embeddings, with no formula for computing the score, no normalization, and no statement of how it depends on r^l_ij, h^l_i, and h^l_j. Because Eq. (12) is the only aggregator that uses attention, and the ablation description for w/o FGL refers to removing an edge-to-node mapping and dot product, the missing definition of alpha-hat is load-bearing for the implementation and for the reported numbers. The exact computation of alpha-hat must be supplied.","section":"IV-D, Eq. (13)"},{"comment":"The complexity analysis is incorrect. The propagation phase in Eq. (9) creates one message per edge, and the aggregation phase in Eq. (10) consumes one message per edge, so the per-layer cost is Theta(e*d) (or Theta(e) with d treated as constant), not O(n). On Freebase, where e is approximately 10^6 and n is approximately 1.8*10^5, this distinction is material. The claimed O(n) total should be replaced with Theta((n+e)*d) per layer or an equivalent statement that makes the dependence on the number of edges explicit.","section":"IV-F"},{"comment":"The summary sentence that EdgeGFL outperforms SeHGNN by an average of 1.51% (micro-F1) and 0.64% (macro-F1) is not supported by the Improve(%) row of Table III. The per-dataset micro-F1 improvements are 0.59, 0.47, 0.27, and 5.38, whose average is 1.68, and the macro-F1 improvements are 0.65, 0.45, 0.27, and 2.53, whose average is 0.98. No simple average of the reported values yields 1.51 and 0.64. Additionally, the text in the same section says Freebase improvements are 5.3% and 0.91%, while Table III reports 5.38% and 2.53%. These numbers should be reconciled and the summary statistic recomputed.","section":"V-D and Table III"},{"comment":"Equation (9) uses the symbol '·' and calls it a dot product, but the mechanism described in Definition 1 and in the Introduction is per-dimension scaling of the source features by the edge preference vector (a Hadamard product). If '·' is an inner product, M^l_ij is a scalar and the message loses the feature dimension that the paper claims to preserve; if it is an elementwise product, the notation is wrong. This needs to be corrected because the entire feature-preference claim rests on per-dimension gating.","section":"IV-C, Eq. (9)"}],"minor_comments":[{"comment":"The dataset bullet descriptions appear mismatched: DBLP is described as a computer science bibliography, ACM as a movie-oriented website, and IMDB as a citation network; the ACM and IMDB descriptions should be swapped.","section":"V-A"},{"comment":"The text gives Freebase improvements of 5.3% and 0.91% in micro-F1 and macro-F1, but Table III reports 5.38% and 2.53%; these numbers should be reconciled.","section":"V-D"},{"comment":"The text 'using a rule or softmax function' should read 'ReLU or softmax function'.","section":"IV-A"},{"comment":"The notation for dimensions is inconsistent: h^l_i is given as an element of R^{1*d} in Table I but as an element of R^{n*d_l} in Section III-A, and W^l is listed with dimensions that do not match its use as a projection matrix in Eqs. (11) and (12).","section":"III-A and Table I"},{"comment":"The citation of CensNet as [35] and translation mechanisms as [36] does not match the reference list, where the CensNet paper is reference [36].","section":"II"},{"comment":"The discussion in the time analysis mentions an 'HGB' model, but no model named HGB appears in the list of compared methods in Section V-B.","section":"V-G"},{"comment":"The axis labels in Figures 6 and 7 contain glyph artifacts (visible as /uni... strings) in the submitted version, making the parameter-sensitivity plots difficult to read.","section":"Figures 6 and 7"}],"recommendation":"major_revision","confidential_remarks":"The primary blockers are the missing definition of alpha-hat and the conflicting model descriptions; these are fixable in a revision but must be addressed before the empirical claims can be evaluated. I would also ask the editor to require the authors to provide the exact aggregation used for each table, the dataset splits, the number of random seeds, and preferably code or a detailed pseudo-code. The manuscript also shows signs of an incomplete editorial pass (dataset bullet mismatches, reference numbering errors, glyph corruption in figures), so a careful revision is needed rather than a minor fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. The core proposal — represent each edge type as a learned vector and use it as a per-dimension multiplicative gate on messages — is simple and reasonable. On the four standard benchmarks the reported margins over Simple-HGN and SeHGNN are small but consistent. The problem is that the manuscript does not actually define the model that produced the numbers. Eqs. (9)–(10) define an edge-gated sum aggregator with no attention, while Eqs. (11)–(13) describe a residual scalar-attention GCN in which the edge vector never appears, and the attention score alpha-hat in Eq. (13) is never specified. Algorithm 1 says use Eqs. (9)–(10); the ablation narrative says the implemented model uses the edge-to-node mapping and dot product. These are different models, and no code is provided. So the state-of-the-art claim is attached to an algorithm that isn't uniquely specified.\n\nWhat's genuinely new is the specific combination of type-based edge embeddings as element-wise gates with L2 normalization and residual connections. That is a close relative of edge-conditioned convolutions and Simple-HGN, but the exact gating formulation is not verbatim in the cited baselines. The paper also does honest ablations: removing the edge gating, L2 norm, nonlinear encoding, and type-based initialization all hurt, and the w/o FGL and w/o EI drops are substantial. If the method is pinned down, the empirical story could be a modest but useful contribution.\n\nBeyond the central under-specification: the complexity analysis in Sec. IV-F is wrong — message passing over e edges is O(e), not O(n); the dataset descriptions in Sec. V-A are scrambled (ACM is a citation network, not a movie site; IMDB is a movie dataset, not a citation network); ref [47] for MH-GCN points to a traffic forecasting paper, not a multiplex heterogeneous GCN; and Eq. (16) is labeled \"with Similar Relation Learning\" though no such term is defined. The reported standard deviations for EdgeGFL are suspiciously tiny (0.01–0.09) compared with the baselines' 0.2–2.5; that should be checked. None of these alone would sink the paper; the undefined alpha-hat does.\n\nWho benefits: readers working on heterogeneous GNNs will get a useful hypothesis about edge-type gating, and a cautionary example of why one must specify the aggregator. As submitted, the paper is not reproducible. I would accept it for peer review in the sense of sending it to a serious referee, because the idea is worth evaluating and the fix is in principle straightforward; but my own verdict on the current version is reject or major revision.","headline":"The edge-gating idea is sensible and the reported gains are plausible, but the paper never specifies which model produced Table III; that is a load-bearing hole.","tokens_in":24946,"tokens_out":2590,"would_cite":false,"duration_ms":24125,"reading_group":"maybe","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 representing edges as learnable multidimensional vectors, used to gate node-feature dimensions during message passing, improves node classification and clustering on heterogeneous graphs, beating the strongest prior…","keywords":["heterogeneous graph neural networks","edge embeddings","feature preference learning","message passing","node classification","node clustering","residual attention","graph representation learning"],"falsifier":"Attempt a faithful reproduction of EdgeGFL on DBLP, ACM, IMDB, and Freebase using only the equations given: the computation of $\\hat{\\alpha}$ in Eq. (13) is unspecified, so no unique implementation exists. If no definition of $\\hat{\\alpha}$ reproduces the reported micro-F1 numbers, in particular 95.98 on DBLP and 69.04 on Freebase, the state-of-the-art claim cannot be substantiated.","tokens_in":1537,"feed_emoji":"🔗","tokens_out":1534,"duration_ms":51582,"temperature":0.7,"pith_summary":"EdgeGFL argues that edge information in graph neural networks should not be a single scalar weight, as in adjacency matrices or attention coefficients, but a multidimensional learnable vector that scales each feature dimension of the messages passing between nodes. The paper claims this per-dimension 'feature preference' lets target nodes keep relevant features and suppress noise, improving node representations on heterogeneous graphs. On four real-world benchmarks, the method reports state-of-the-art node classification and clustering results, beating the strongest prior heterogeneous GNN, SeHGNN, by 1.51% micro-F1 and 0.64% macro-F1 on average, and by more than 5% micro-F1 on Freebase. If true, this places edge embeddings at the center of message passing rather than treating them as an optional auxiliary signal.","feed_headline":"EdgeGFL tops SeHGNN by 1.51% micro-F1 across four graphs","feed_subtitle":"Per-dimension edge gating gives nodes finer information in heterogeneous graphs.","key_machinery":"The central object is the edge feature preference vector $r^l_{ij}$: for each edge type, an initialized encoding vector is mapped into the node feature space by a shared learnable transformation, and this vector multiplies the source node representation elementwise in Eq. (9), so each feature dimension of a message is scaled differently. This replaces the scalar aggregation weights of GCN and GAT with a multidimensional gate. The aggregation phase then uses sum aggregators with node and edge residual connections, Eq. (11)-(13), where the residual attention coefficient $\\hat{\\alpha}$ is said to combine edge type embeddings and node embeddings, although its concrete formula is not given.","core_discovery":"The central claim is that enriching edges from scalar weights to multidimensional, learnable relation vectors makes message passing selective at the feature-dimension level, so that each neighbor's contribution is amplified or dampened per feature rather than scaled uniformly. The paper proposes EdgeGFL, a message-passing framework where each edge type is mapped to an initialization vector, transformed into the node feature space, and multiplied elementwise with the source node's representation in Eq. (9). This refined message is then aggregated with residual connections on both node representations and attention scores, and the final embeddings are L2-normalized. On the DBLP, ACM, IMDB, and Freebase heterogeneous graphs, the paper reports improvements over SeHGNN averaging 1.51% in micro-F1 and 0.64% in macro-F1 for node classification, with a 5.38% micro-F1 gain on Freebase, and consistent gains in clustering ARI and NMI.","pith_inferences":["A natural testable extension is to replace the edge-type dictionary initialization with continuous relation features, such as text or attribute embeddings of relations, which the paper does not explore.","The missing definition of $\\hat{\\alpha}$ in Eq. (13) means the current write-up does not provide a fully reproducible algorithm; a reader cannot tell whether the reported gains come from the edge-vector gating itself or from an unspecified attention module.","If the gains hold, the per-dimension gating mechanism could serve as a generic relational inductive bias for message passing, applicable beyond heterogeneous graphs to any network with typed or attributed edges."],"forward_implications":["Heterogeneous graph models may not need hand-designed meta-path selection, since learned edge vectors carry relation semantics; EdgeGFL reports strong results without meta-path enumeration.","Per-dimension edge gating gives larger relative gains on Freebase, which has 36 edge types, suggesting the benefit grows with the number of relation types.","Deep GCN stacks with node and edge residual connections can avoid the over-smoothing that limits ordinary convolutional layers; EdgeGFL reports best performance near two layers and stable convergence within about 60 training rounds.","L2 normalization of the output embeddings makes the learned representations directly usable for retrieval-style tasks such as clustering and similarity search.","Replacing scalar edge weights with vectorized edge features could transfer to any graph where relation types carry rich semantics, including knowledge graphs and recommendation networks."],"supporting_citations":[{"why":"Provides the state-of-the-art SeHGNN baseline that EdgeGFL claims to beat by 1.51% micro-F1 and 0.64% macro-F1 on average.","marker":"[46]"},{"why":"Supplies the closest competing design, Simple-HGN, with learnable edge type embeddings, residual connectivity, and L2 normalization, which EdgeGFL extends and compares against.","marker":"[45]"},{"why":"Defines the GCN message-passing framework whose scalar adjacency-weight aggregation EdgeGFL generalizes to vectorized edge preferences.","marker":"[13]"},{"why":"Defines attention-based scalar weights in GAT, the approach EdgeGFL replaces with per-dimension feature preference.","marker":"[7]"},{"why":"Motivates applying residual connections to attention scores, used in Eq. (13) for the edge residual.","marker":"[38]"},{"why":"Provides the Freebase knowledge-graph data used in the experiments, where EdgeGFL reports its largest improvement.","marker":"[39]"},{"why":"Supplies the survey-and-benchmark protocol that guides the Freebase subgraph extraction and evaluation setting.","marker":"[40]"},{"why":"Defines the HAN heterogeneous graph attention baseline with node-level and semantic-level attention, which EdgeGFL compares against.","marker":"[10]"}],"fun_headline_variants":["EdgeGFL: Edge vectors, not scalars, boost GNN node classification","Per-dimension edge gating lifts GNN micro-F1 by 1.51%","EdgeGFL learns edge embeddings to refine node features","EdgeGFL beats SeHGNN by 1.51% micro-F1 on heterogeneous graphs","Multidimensional edge features sharpen heterogeneous graph embeddings"],"cache_read_input_tokens":27008,"weakest_assumption_plain":"The reported results presuppose a concrete formula for the residual attention coefficient $\\hat{\\alpha}$ in Eqs. (11)-(13), but that formula, which combines edge type embeddings and node embeddings, is never defined anywhere in the paper.","fun_headline_variants_meta":{"raw":{"variants":["EdgeGFL: Edge vectors, not scalars, boost GNN node classification","Per-dimension edge gating lifts GNN micro-F1 by 1.51%","EdgeGFL learns edge embeddings to refine node features","EdgeGFL beats SeHGNN by 1.51% micro-F1 on heterogeneous graphs","Multidimensional edge features sharpen heterogeneous graph embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3339,"prompt_tokens":946,"completion_tokens":2393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2296}},"tokens_in":562,"tokens_out":2393,"duration_ms":16730,"temperature":1.0,"reasoning_tokens":2296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:36:34.309366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Attempt a faithful reproduction of EdgeGFL on DBLP, ACM, IMDB, and Freebase using only the equations given: the computation of $\\hat{\\alpha}$ in Eq. (13) is unspecified, so no unique implementation exists. If no definition of $\\hat{\\alpha}$ reproduces the reported micro-F1 numbers, in particular 95.98 on DBLP and 69.04 on Freebase, the state-of-the-art claim cannot be substantiated.","supporting_citations":[{"cited_title":"Simple and efficient heterogeneous graph neural network,","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art SeHGNN baseline that EdgeGFL claims to beat by 1.51% micro-F1 and 0.64% macro-F1 on average."},{"cited_title":"Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the closest competing design, Simple-HGN, with learnable edge type embeddings, residual connectivity, and L2 normalization, which EdgeGFL extends and compares against."},{"cited_title":"RealFormer: Transformer Likes Residual Attention","cited_arxiv_id":"2012.11747","evidence_quote":"Motivates applying residual connections to attention scores, used in Eq. (13) for the edge residual."},{"cited_title":"Freebase: a collaboratively created graph database for structuring human knowledge,","cited_arxiv_id":null,"evidence_quote":"Provides the Freebase knowledge-graph data used in the experiments, where EdgeGFL reports its largest improvement."},{"cited_title":"Heterogeneous network representation learning: A unified framework with survey and benchmark,","cited_arxiv_id":null,"evidence_quote":"Supplies the survey-and-benchmark protocol that guides the Freebase subgraph extraction and evaluation setting."},{"cited_title":"Heterogeneous graph attention network,","cited_arxiv_id":null,"evidence_quote":"Defines the HAN heterogeneous graph attention baseline with node-level and semantic-level attention, which EdgeGFL compares against."}],"review_version":1}