{"id":"3a9f9d4a-44da-4fa3-bc3d-d2d0171ae1e3","arxiv_id":"2506.16754","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MHCL learns per-metapath hyperbolic spaces with learnable curvature and a hyperbolic contrastive loss to separate metapath embeddings, reporting modest gains over existing heterogeneous graph embedding baselines.","lead":"This paper introduces MHCL, a graph-learning model that represents each metapath in a heterogeneous graph in its own hyperbolic space with a learnable curvature, and uses a contrastive loss to keep metapath embeddings distinct. It reports gains over prior heterogeneous graph embedding methods on four benchmark datasets, but the contrastive loss as written would push positive pairs apart.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (10) as written has no negative logarithm, so minimizing it repels positive metapath samples and attracts negatives, which is the opposite of the claimed contrastive mechanism. Without code or a corrected loss, the central contribution is unverified.","rationale":"I read the manuscript in good faith and identified the same load-bearing weakness as the reader: the mathematical form of the contrastive loss in Eq. (10) is inconsistent with the mechanism described in the text. The paper claims that the main contribution is enhancing the discriminability of metapath embeddings by minimizing same-metapath distance and maximizing different-metapath distance in hyperbolic space. Eq. (10) as printed cannot do this; its gradient is opposite. This is not merely a stylistic issue because Algorithm 1 and Eq. (18) use L_hyp directly in the total loss, so the actual training signal depends on the exact expression. I also checked whether other parts of the paper provide independent support: the ablations and t-SNE visualizations do support the narrative, but they only show what the authors' implementation did, not whether it matches Eq. (10). The lack of released code leaves the discrepancy unresolved. I do not treat this as evidence of misconduct; the most plausible reading is that the intended objective is the negative log of a contrastive ratio and the printed equation omitted the log. Under that reading the central idea is coherent and the empirical results are plausible, so the appropriate final verdict remains CONDITIONAL, not REJECT. Since the reader already assigned CONDITIONAL, I recommend UNCHANGED. A concrete comparison between the printed loss and the corrected -log objective would settle whether the concern is a typo or a real flaw.","tokens_in":21103,"tokens_out":4457,"duration_ms":44329,"concrete_test":"Request the official code or re-implement the model from the paper, then run the exact objective of Eq. (10) and the corrected InfoNCE form L = -log[ exp(-d_pos/tau) / ( exp(-d_pos/tau) + sum_j exp(-d_neg_j/tau) ) ] on IMDB and DBLP with the paper's hyperparameters. If minimizing Eq. (10) as written degrades or anti-separates metapath embeddings while the corrected -log form reproduces Table III and Figure 8, then Eq. (10) is a typo and the claim survives conditionally; if the released code matches Eq. (10) and still outperforms baselines, the reported mechanism requires a different explanation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's novelty is the metapath-based hyperbolic contrastive loss, defined in Eq. (10) as L_hyp = sum_phi exp(-d_pos/tau) / sum_j exp(-d_neg_j/tau), where d_pos is the distance to the positive sample and d_neg_j are distances to negative samples. This is a ratio with no negative logarithm and no positive term in the denominator. Minimizing it has the wrong gradient: the loss decreases as the positive distance d_pos increases, because the numerator shrinks, and it also decreases as negative distances d_neg_j decrease, because the denominator grows. Thus gradient descent on Eq. (10) as written maximizes distance to the positive sample and minimizes distance to negative samples, directly contradicting Section IV-D.3, which states that the method minimizes distance between embeddings of the same metapath and maximizes distance between embeddings of different metapaths. Algorithm 1 line 16 says to calculate L_hyp in (10), and the total loss (18) adds lambda times L_hyp, so if the implemented loss matches the printed equation, the contrastive component is anti-contrastive. The ablation results in Table VI and the t-SNE visualizations in Figure 8 are consistent with a corrected InfoNCE-style objective, but the paper provides no code and no derivation of the implemented objective. Because the central claimed improvement over MSGAT rests entirely on this loss, the exact form of the training objective is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MHCL, a heterogeneous graph embedding framework that uses multiple metapath-specific hyperbolic spaces with learnable curvatures to capture diverse power-law structures, plus a metapath-based hyperbolic contrastive loss intended to separate embeddings from different metapaths. The method is evaluated on node classification, clustering, and link prediction across IMDB, DBLP, ACM, LastFM, and a synthetic graph, with ablations, robustness tests, hyperparameter sensitivity, complexity experiments, and t-SNE visualizations. The central claimed contribution is the contrastive loss in Eq. (10), which the text states should minimize the distance between same-metapath embeddings and maximize the distance between different-metapath embeddings.","tokens_in":21480,"tokens_out":4080,"duration_ms":45695,"significance":"If the method works as described, the multi-hyperbolic-space architecture with per-metapath learnable curvatures is a meaningful extension of prior hyperbolic heterogeneous GNNs such as MSGAT, and the empirical study is reasonably broad: four real-world datasets plus a synthetic robustness study, multiple downstream tasks, ablations, and complexity measurements. The paper does not provide code, and the central contrastive-loss equation is not a valid contrastive objective as printed. The positive experimental results and the t-SNE visualizations are consistent with a corrected InfoNCE-style loss, but the exact implemented objective must be stated before the claimed improvement over MSGAT can be evaluated. No circularity issue arises from the use of the authors' earlier HHGAT and MSGAT as baselines, since they are trained and evaluated on held-out tasks.","major_comments":[{"comment":"The printed contrastive loss L_hyp is not a valid contrastive objective. As written, L_hyp = sum_phi exp(-d_c(hat_h^phi_v, p(h^phi_v))/tau) / sum_{j(phi) in Phi \\ {phi}} exp(-d_c(hat_h^phi_v, hat_h^{j(phi)}_v)/tau) has no negative logarithm. Minimizing this ratio decreases the numerator when the positive distance d_c(hat_h^phi_v, p(h^phi_v)) increases, and decreases the denominator when negative distances decrease, so gradient descent on Eq. (10) as printed would separate an anchor from its positive sample and pull it toward negative samples. This directly contradicts the stated goal in Section IV-D.3 and the abstract. Since Eq. (18) adds lambda * L_hyp to the task loss, the printed training objective is anti-contrastive. Please correct Eq. (10) to the negative log of the ratio (or an equivalent InfoNCE form), update Algorithm 1 and all references to it, and state whether the implemented loss matches the corrected equation. Because no code is provided, the current text does not allow the reader to verify which objective was actually optimized.","section":"Section IV-D.3, Eq. (10); Algorithm 1 line 16; Eq. (18)"},{"comment":"The formula in Eq. (10) also contains an inconsistent use of the temperature tau: in the numerator tau appears inside the distance as p(h^phi_v)/tau, while in the denominator it scales the distance, i.e., d_c(hat_h^phi_v, hat_h^{j(phi)}_v)/tau. Additionally, the loss is written for a single node v with no aggregation over the target node set V_t, although Algorithm 1 takes a single node v as input. Please specify how L_hyp is aggregated over nodes (e.g., summed or averaged over V_t) and correct the placement of tau. These details are needed to reproduce the reported loss values and the experimental results.","section":"Section IV-D.3, Eq. (10); Algorithm 1"},{"comment":"The claim that the overall time complexity increases linearly with respect to the maximum metapath length l is not supported by the stated complexity analysis. Metapath instance sampling via breadth-first search from a node with branching factor b and maximum length l enumerates O(b^l) instances in the worst case (up to type constraints), not O(l). The empirical near-linear growth in Fig. 9 may reflect the specific degree distributions of the ACM dataset, but the text should restate the worst-case complexity in terms of the branching factor and explain why the observed training time grows approximately linearly on this dataset. As written, the complexity analysis in Section IV-G is inaccurate.","section":"Section IV-G and Section V-K, Fig. 9"}],"minor_comments":[{"comment":"There are several typos and wording issues, including 'postive' and 'metpath-specific' in Section IV-D, 'Additonally' in Section V-K, and 'metapfath' in the Introduction. A careful proofread is needed.","section":"Throughout"},{"comment":"The hyperbolic distance formula in Eq. (11) uses 1/sqrt(|c|) although c is already defined to be positive in Definition 4; this is needlessly confusing and can be written as 1/sqrt(c).","section":"Section III-B, Eq. (11)"},{"comment":"The alignment step in Eq. (8) applies a linear transformation W2 in the tangent space at the origin, but the text does not specify whether this tangent-space coordinate representation is shared across different curvatures; a brief clarification of the coordinate convention would improve reproducibility.","section":"Section IV-D.1, Eq. (8)"},{"comment":"The main result tables are extremely wide and difficult to read in the provided format; splitting them by dataset or grouping competitors differently would improve readability.","section":"Tables III-V"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is Eq. (10): as printed, the contrastive loss is anti-contrastive, and this equation is the core claimed novelty. The fix is local in scope (add the negative logarithm and correct the temperature placement), but without code or a corrected equation the experimental results cannot be attributed to the claimed mechanism. I would like the editor to ask the authors for the exact implemented loss and, ideally, code release. The paper also has no code and the complexity analysis in Section IV-G is incorrect as stated, so at least one round of revision with careful re-verification is necessary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. This paper extends MSGAT—the authors' own ICDM 2024 model—with a metapath-based hyperbolic contrastive loss, and reports consistent 1–3 point gains over strong baselines on four datasets. That is a useful increment for the heterogeneous graph embedding subfield. But the central equation, Eq. (10), is written in a way that does the opposite of what the text claims. Minimizing L_hyp as printed shrinks the numerator by pushing the positive distance up and grows the denominator by pulling negative distances down. Section IV-D.3 says same-metapath embeddings should be pulled together and different-metapath embeddings pushed apart; the equation drives the reverse. That is not a nitpick—the entire novelty claim rests on that loss, and Algorithm 1 plus Eq. (18) include it directly. No code is provided, so we cannot tell whether the implemented loss is the intended InfoNCE-style objective with a negative log. The ablation results and t-SNE figures are consistent with a corrected loss, but as printed the mechanism is unverified.\n\nWhat is genuinely decent: the multi-hyperbolic-space architecture with per-metapath learnable curvatures is a reasonable idea, and the paper is honest that MHCL Single and MHCL w/o cont reproduce their own earlier HHGAT and MSGAT. The experiments are broad: four real datasets, node classification, clustering, link prediction, ablations, synthetic robustness, hyperparameter sensitivity, and complexity measurements. The baselines are appropriate, including McH-HGCN and SHAN. Self-citation is explicit and not a problem here.\n\nSoft spots besides the equation: the motivation for per-metapath curvature is asserted rather than demonstrated. Figure 2 shows different Gromov deltas per metapath, but there is no analysis connecting learned curvatures to those deltas, or showing that separate curvatures are needed once the contrastive loss is in place. The gains over MSGAT are modest, which is acceptable, but the contribution margin is thin. The lack of code is a real problem for an empirical systems paper, especially when the loss is mis-specified.\n\nWho is this for? Researchers working on heterogeneous graph embedding who want a benchmarkable variant of multi-curvature hyperbolic GNNs. The architecture is worth knowing about, but only after the authors fix the loss equation and release code. If the implemented loss is a standard InfoNCE variant, the empirical story may hold up. As printed, the paper is not ready.\n\nMy recommendation: send to peer review, but with a strong request that the authors correct Eq. (10), specify the exact objective (negative log, positive included in the denominator), and release code. The idea is plausible and the experimental scope merits referee time; the flaw is central but likely fixable.","headline":"A modest extension of MSGAT whose central contrastive loss (Eq. 10) is mis-specified to repel positives and attract negatives — with no code, the reported gains cannot be credited to the proposed method.","tokens_in":22028,"tokens_out":4781,"would_cite":false,"duration_ms":43550,"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":"This paper claims that assigning each metapath of a heterogeneous graph its own hyperbolic space with a learned curvature, and then applying a contrastive loss that separates embeddings of different metapaths, yields better node…","keywords":["heterogeneous graph embedding","hyperbolic space","metapath","contrastive learning","graph neural network","power-law structure","Poincaré ball","node classification"],"falsifier":"Training MHCL with the exact loss of Eq. (10) and observing that same-metapath embeddings grow more distant and different-metapath embeddings more similar would show the described mechanism is not what the formula implements; alternatively, zeroing the contrastive term and finding no significant drop on the reported datasets would falsify the claimed benefit of contrastive separation.","tokens_in":20884,"feed_emoji":"🕸️","tokens_out":7239,"duration_ms":63418,"temperature":0.7,"pith_summary":"MHCL is designed around a simple premise: a heterogeneous graph contains several different power-law structures, one per metapath, and a single hyperbolic space cannot faithfully encode all of them at once. The paper proposes to give each metapath its own Poincaré ball with a learnable negative curvature, aggregate metapath instances inside each ball with attention, and then align all metapath embeddings in a unified hyperbolic space where a contrastive loss keeps different metapaths apart. On the IMDB, DBLP, ACM, and LastFM datasets, the paper reports that MHCL outperforms state-of-the-art Euclidean and hyperbolic heterogeneous GNNs on node classification, clustering, and link prediction, and its ablations show that both the multi-space geometry and the contrastive term contribute to the gains. The significance, if the results hold, is that geometry should be matched to each semantic relation rather than shared across the whole graph.","feed_headline":"Per-metapath curvature plus contrastive loss beats single-space GNNs","feed_subtitle":"Metapath-specific hyperbolic geometry and contrastive separation lift node, clustering, and link prediction on four benchmarks.","key_machinery":"The load-bearing object is a family of metapath-specific Poincaré balls $\\mathbb{D}^{d,c_\\phi}$, one per metapath $\\phi$, each with its own learnable negative curvature $-c_\\phi$ ($c_\\phi > 0$). Inside each ball, metapath instances are aggregated by intra-hyperbolic attention (Eqs. 4-6) with multi-head extension (Eq. 7); embeddings are then aligned into a unified hyperbolic space via the shared tangent space at the origin (Eq. 8) and combined by inter-hyperbolic attention (Eqs. 12-14). The contrastive loss $L_{\\mathrm{hyp}}$ (Eq. 10), using hyperbolic distance $d_c$ (Eq. 11), is what the paper claims enforces discriminability of metapath embeddings.","core_discovery":"The central discovery claimed is that the power-law 'shape' of a heterogeneous graph is not a single global property but is metapath-specific, so an embedding model should use multiple hyperbolic geometries whose curvatures adapt per metapath. MHCL samples metapath instances via BFS up to a maximum length, embeds them in metapath-specific Poincaré balls, aggregates them with parameterized attention to form metapath embeddings, maps those into a unified hyperbolic space through the shared tangent space at the origin, and then applies a hyperbolic contrastive loss so that embeddings of different metapaths are pushed apart while each metapath stays consistent with its own embedding. The paper presents experiments on four real-world datasets and a synthetic graph, reporting consistent outperformance over 13 baselines, plus ablation and visualization evidence that the contrastive loss separates metapath clusters.","pith_inferences":["A natural next step the paper does not pursue: use the per-metapath Gromov $\\delta$-hyperbolicity values (which the paper computes for motivating examples) to initialize or regularize the learned curvatures, potentially improving convergence and stability.","A testable claim implicit in the multi-space premise: on a heterogeneous graph whose metapaths all induce nearly identical power-law distributions (similar Gromov deltas), MHCL should converge to a single effective curvature; this could be checked on synthetic graphs with controlled metapath distributions.","The contrastive separation of metapath embeddings is measured only in the unified space; one could probe whether the learned per-metapath curvatures themselves encode semantic differences, for example by inspecting curvature values per metapath across datasets."],"forward_implications":["If MHCL's claim is right, heterogeneous graph models should treat curvature as a per-metapath parameter rather than a global constant, since different metapaths induce measurably different power-law distributions.","The contrastive separation of metapath embeddings should transfer to any downstream task that relies on distinguishing semantic relations, including recommendation and knowledge-graph completion, not just the four datasets tested.","Because MHCL samples metapaths with BFS up to a maximum length instead of relying on hand-crafted metapath definitions, the approach removes the need for domain-specific metapath engineering while keeping linear time complexity in the maximum metapath length.","The paper's ablations indicate that geometric capacity and contrastive regularization contribute independently, since removing either component degrades performance across all tested datasets."],"supporting_citations":[{"why":"the multi-hyperbolic-space predecessor without contrastive learning; ablation comparison isolates the effect of the contrastive loss.","marker":"[27]"},{"why":"the single-hyperbolic-space metapath model; ablation comparison isolates the effect of using multiple spaces.","marker":"[15]"},{"why":"the foundational metapath-based heterogeneous GNN whose node-level and semantic-level attention structure MHCL adapts.","marker":"[2]"},{"why":"a multi-curvature hyperbolic heterogeneous GNN baseline that assigns curvatures per link type.","marker":"[26]"},{"why":"the hyperbolic contrastive learning baseline for homogeneous graphs that the paper compares against.","marker":"[29]"},{"why":"the hyperbolic GNN baseline for homogeneous graphs representing single-space hyperbolic geometry.","marker":"[34]"},{"why":"a hyperbolic heterogeneous GNN baseline learning power-law structures from simplicial complexes.","marker":"[14]"},{"why":"a metapath aggregation baseline that incorporates intermediate nodes, used as a strong Euclidean heterogeneous GNN.","marker":"[24]"}],"fun_headline_variants":["Per-metapath hyperbolic geometry plus contrastive loss beats single-space","Diverse hyperbolic curvatures per metapath, separated by contrastive loss","Multiple hyperbolic spaces adapt to each metapath structure","Metapath-specific hyperbolic embeddings, made separable via contrastive loss","Contrastive learning in per-metapath hyperbolic spaces outperforms baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's described contrastive mechanism depends on the implemented loss being the negative logarithm of the ratio in Eq. (10); as written, minimizing that equation would push positive metapath pairs apart instead of together, and the paper does not supply the form actually implemented.","fun_headline_variants_meta":{"raw":{"variants":["Per-metapath hyperbolic geometry plus contrastive loss beats single-space","Diverse hyperbolic curvatures per metapath, separated by contrastive loss","Multiple hyperbolic spaces adapt to each metapath structure","Metapath-specific hyperbolic embeddings, made separable via contrastive loss","Contrastive learning in per-metapath hyperbolic spaces outperforms baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3069,"prompt_tokens":965,"completion_tokens":2104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2012}},"tokens_in":581,"tokens_out":2104,"duration_ms":15654,"temperature":1.0,"reasoning_tokens":2012,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:19:07.958937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Training MHCL with the exact loss of Eq. (10) and observing that same-metapath embeddings grow more distant and different-metapath embeddings more similar would show the described mechanism is not what the formula implements; alternatively, zeroing the contrastive term and finding no significant drop on the reported datasets would falsify the claimed benefit of contrastive separation.","supporting_citations":[{"cited_title":"Multi-hyperbolic space-based heterogeneous graph attention network,","cited_arxiv_id":null,"evidence_quote":"the multi-hyperbolic-space predecessor without contrastive learning; ablation comparison isolates the effect of the contrastive loss."},{"cited_title":"Hyperbolic heterogeneous graph attention networks,","cited_arxiv_id":null,"evidence_quote":"the single-hyperbolic-space metapath model; ablation comparison isolates the effect of using multiple spaces."},{"cited_title":"Mch-hgcn: Multi-curvature hyperbolic heteroge- neous graph convolutional network with type triplets,","cited_arxiv_id":null,"evidence_quote":"a multi-curvature hyperbolic heterogeneous GNN baseline that assigns curvatures per link type."},{"cited_title":"Hyperbolic graph convolu- tional neural networks,","cited_arxiv_id":null,"evidence_quote":"the hyperbolic GNN baseline for homogeneous graphs representing single-space hyperbolic geometry."},{"cited_title":"Multi-order relations hyperbolic fusion for heterogeneous graphs,","cited_arxiv_id":null,"evidence_quote":"a hyperbolic heterogeneous GNN baseline learning power-law structures from simplicial complexes."},{"cited_title":"Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding,","cited_arxiv_id":null,"evidence_quote":"a metapath aggregation baseline that incorporates intermediate nodes, used as a strong Euclidean heterogeneous GNN."}],"review_version":2}