{"id":"8719a894-5e23-4b3e-9a92-6a9c9bdde8f2","arxiv_id":"2504.14205","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DHMP separates fraud graphs into homophilic and heterophilic channels with a learned edge splitter and reports improved fraud detection on YelpChi, Amazon, and FDCompCN.","lead":"Dual-channel Heterophilic Message Passing (DHMP) splits a fraud graph into homophilic and heterophilic subgraphs using a label-trained edge classifier, then propagates messages separately through complementary filters before fusing them. On three public fraud datasets the authors report small but consistent gains over several graph fraud detection baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical superiority is not established: single-run results, no error bars, small AUC margins, and an internal GMean discrepancy leave the central outperformance claim unprotected.","rationale":"The reader's CONDITIONAL verdict already captures the need for robustness reporting, error bars, and code verification, and I do not see a reason to move away from CONDITIONAL. However, I locate the single most load-bearing weakness in the empirical evidence itself rather than in the heterophily separator. Separator accuracy is important for the mechanistic story: if edge types are misclassified, the two channels are contaminated and the claimed frequency separation is degraded. But the central claim is comparative and empirical, so a direct test of separator accuracy would not by itself decide whether DHMP beats the baselines. Multi-seed reruns with significance testing would settle the central claim directly. The internal inconsistency between Tables II and III, and the YelpChi Recall reversal against H2-FDetector, make the single-run table especially fragile. Thus the appropriate recommendation is to keep the verdict conditional, with the concrete multi-seed check as the binding condition.","tokens_in":97835,"tokens_out":7258,"duration_ms":70656,"concrete_test":"Reproduce Table II on all three datasets with DHMP and at least the top three baselines (GFAN, BWGNN, H2-FDetector) using 10 independent seeds or random splits under the paper's own protocol; report mean±std for Recall, F1-macro, AUC, and GMean, plus a paired bootstrap or Wilcoxon test of DHMP versus each baseline. If the Amazon or FDCompCN AUC advantage shrinks below noise or fails significance, the central outperformance claim is not established. Also re-run the exact DHMP configuration to reconcile the YelpChi GMean discrepancy between Table II (83.68) and Table III (84.38).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's 'DHMP outperforms existing methods.' The evidence is Table II: one number per method per metric, with no standard deviations, no number of seeds, and no significance tests. Several reported margins over the strongest baseline are small: Amazon AUC 92.32 vs 91.91 (0.41 pp), FDCompCN AUC 61.43 vs 60.79 (0.64 pp), and YelpChi AUC 91.80 vs 90.69 (1.11 pp). At this scale, run-to-run variation from split or initialization could easily flip the ranking. This fragility is concrete: Table II reports DHMP YelpChi GMean as 83.68, while Table III reports 84.38 for the same model, indicating instability or a transcription error. The claim of higher Recall than all baselines also already fails in Table II: H2-FDetector reaches 84.61 on YelpChi vs DHMP's 84.39. Without variance or paired tests, the positive margins are indistinguishable from noise. This is load-bearing because if reruns do not reproduce the margins, the paper's headline conclusion is unsupported even if the architecture is novel.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DHMP, a GNN-based fraud detection framework that splits the graph into homophilic and heterophilic subgraphs via a learned MLP separator, propagates messages in two channels using a shared filter Wf and separate residual transforms, fuses the channel embeddings and multi-relation representations, and trains with a weighted classification plus heterophily-separation loss. Experiments on YelpChi, Amazon, and FDCompCN compare DHMP against thirteen baselines on Recall, F1-macro, AUC, and GMean, with ablations, sensitivity analyses, and a visualization study.","tokens_in":98089,"tokens_out":4521,"duration_ms":41445,"significance":"The proposed architecture is plausible and addresses a real challenge in fraud detection: the coexistence of homophily and heterophily in fraud graphs. Treating heterophilic edges as informative signals rather than pruning them is a reasonable direction, and the multi-channel design with residual re-scaling is clearly described. The paper uses three standard public datasets, compares against thirteen baselines, and provides ablation and sensitivity experiments, which are positive elements. If the reported performance is reproducible, the gains over strong baselines on FDCompCN and Amazon would be meaningful. However, the empirical superiority claim is not yet established because the results appear to be single-run, lack variance or significance tests, and contain an internal numerical inconsistency that weakens confidence in the reported numbers.","major_comments":[{"comment":"Table II reports DHMP YelpChi GMean as 83.68, while Table III reports 84.38 for the same model in the 'DHMP' row. This internal inconsistency must be resolved and the correct value stated. More generally, all results in Tables II and III appear to be single runs: no standard deviations, no number of seeds, and no significance tests are reported. The margins over the strongest baselines are small (e.g., Amazon AUC 92.32 vs 91.96; FDCompCN AUC 61.43 vs 60.79), so without run-to-run variability the central claim that 'DHMP outperforms existing methods' is not statistically supported. Please provide means and standard deviations over multiple random seeds and pairwise significance tests (or confidence intervals) for the main comparisons, and clarify the GMean discrepancy.","section":"IV-B, Tables II and III"},{"comment":"The text states that compared to GNN-based fraud detection approaches, DHMP improves Recall on YelpChi by at least 1.97%. This is contradicted by Table II itself: H2-FDetector achieves Recall 84.61 while DHMP achieves 84.39. The same table also shows that the GMean margin over GFAN is only 0.02 points (83.68 vs 83.66). The claim of consistent improvement across all metrics must be corrected, or the experimental protocol changed so that the stated conclusion matches the reported data.","section":"IV-B, text after Table II"},{"comment":"The heterophily separator is the component that determines which edges are routed to the homophilic and heterophilic propagation channels, and the auxiliary loss LH only supervises edges in the training set. The paper provides no evaluation of separator accuracy on validation or test edges, and no analysis of how misclassified edges affect the downstream channel representations. Because this module is load-bearing for the entire architecture, please report the separator's classification performance (e.g., accuracy and F1 on edge labels, broken down by homophilic vs heterophilic), and ideally include an oracle-edge comparison to quantify the impact of separator errors on the full model.","section":"III-B, Heterophily Separation"}],"minor_comments":[{"comment":"RQ3 and RQ4 are identical in the list: both read 'What is the influence of differing model parameters on the operational effectiveness of DHMP?' RQ4 should refer to the visualization study.","section":"IV, research questions"},{"comment":"Algorithm 1 has notation errors: the homophilic propagation loop updates 'z−u' but should update 'z+u', and the heterophilic loop also assigns 'z−u' with Eq. 8 to Eq. 10. Please correct the superscripts and the equation ranges.","section":"Algorithm 1"},{"comment":"After Eq. (9), the text says 'where bf, Wn and bn2 are learnable parameters', but bf is not used in Eq. (8) or Eq. (9); this is likely a typo for bn1 or an undefined variable and should be fixed.","section":"III-C, Eq. (9)"},{"comment":"The numbers in Tables II and III contain spacing artifacts such as '74 .35', '91 .80', and '84 .38'. These should be formatted consistently.","section":"Tables II and III"},{"comment":"The text in Section IV-E refers to 'DMPH' in the sentence 'It is worth noting that DMPH can significantly compress...'; this should be 'DHMP'.","section":"IV-E, Visualization"},{"comment":"The abstract says 'shared weights to capture signals at different frequencies independently', but the implementation shares Wf while using separate Wp and Wn for the two channels; the wording 'shared weights' and 'independently' should be reconciled or clarified.","section":"Abstract and III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes an interesting architecture and the topic is within the journal's scope, but the experimental evidence as reported is too weak to support the headline claim. The internal GMean inconsistency and the contradicted Recall claim suggest that the results should be re-verified carefully by the authors before publication. Given the small margins over strong baselines, the editor may wish to require multi-seed results with standard deviations, significance tests, and possibly a code/data release to verify the numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"X—here's the short version. The DHMP architecture is a reasonable synthesis of things already in the literature: a learned edge splitter, paired low/high-pass filters (like FAGCN/BWGNN), separate homophilic and heterophilic propagation (like H2-FDetector), and multi-relation fusion. What is new is the specific combination, including the shared Wf and I−Wf filter pair and the degree-rescaled residual. That is a legitimate architecture contribution, and the ablations indicate each piece adds something. The paper is clearly written and the authors tried to engage with existing methods.\n\nThe problem is the evidence for 'DHMP outperforms existing methods.' Table II has one number per method per metric, no standard deviations, no number of seeds, no significance tests. Several margins are small: Amazon AUC 92.32 vs 91.91, FDCompCN 61.43 vs 60.79. At that scale, run-to-run variation could easily flip the ranking. More concretely, the paper claims DHMP has higher Recall than all baselines, but its own Table II shows H2-FDetector at 84.61 vs DHMP's 84.39 on YelpChi. And the GMean for DHMP on YelpChi is 83.68 in Table II but 84.38 in Table III—same model, same dataset, different numbers. That could just be a typo, but with no code verification it leaves the central claim unprotected.\n\nThere is also a structural assumption that could use analysis: the edge-type classifier is trained on training node labels and then applied to all edges, but there is no measurement of its accuracy or how misclassifications propagate into the two channels. That is a real gap, though it is fixable and does not invalidate the idea.\n\nThe duplicate RQ3/RQ4 is a minor editorial slip. The missing variance and the internal inconsistency are the substantive issues.\n\nWho this is for: the GNN fraud detection subfield, especially people working on heterophily and camouflage. They will want to know whether the edge-splitting idea actually generalizes. With added variance reporting and a corrected table, this could be a solid empirical paper. I would send it to peer review—it deserves referee time—but reviewers should insist on multiple runs, significance tests, and code verification. I would not cite it yet.","headline":"A sensible new combination of existing ideas for fraud detection, but the outperformance claim needs variance reporting and a fix to an internal inconsistency before it is credible.","tokens_in":98624,"tokens_out":1997,"would_cite":false,"duration_ms":20315,"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":"Splitting a fraud graph into same-label and cross-label edges before message passing, then propagating through twin channels with complementary filters, is what lets a GNN outperform thirteen baselines on three fraud datasets.","keywords":["fraud detection","graph neural networks","heterophily","message passing","homophily","label imbalance","multi-relational graph","node classification"],"falsifier":"Run DHMP with the same settings but replace predicted edge types with ground-truth edge types on validation and test edges; if AUC does not rise, the separator is not the source of the gain. Conversely, shuffle edge-type labels during training while keeping the same loss; if performance holds, the dual-channel separation is not doing the claimed work. Either experiment would settle whether the mechanism is load-bearing.","tokens_in":97672,"feed_emoji":"🕵️","tokens_out":4445,"duration_ms":41902,"temperature":0.7,"pith_summary":"This paper tries to establish that fraud detection on graphs improves when a message-passing network stops smoothing away cross-label edges. It claims that splitting a fraud graph into a homophilic subgraph (edges linking similar nodes) and a heterophilic subgraph (edges linking dissimilar nodes), then propagating through twin channels with shared weights, lets a model capture both low-frequency and high-frequency signals. The paper reports that on YelpChi, Amazon, and FDCompCN this dual-channel design outperforms thirteen baselines on Recall, F1-macro, AUC, and GMean. The practical stake is that fraudsters deliberately connect to benign users to camouflage themselves, so a detector that can use those deceptive edges as a signal instead of discarding them should be more robust.","feed_headline":"Splitting fraud graphs by edge type tops 13 baselines","feed_subtitle":"A dual-channel GNN keeps same-label and cross-label messages separate, lifting fraud AUC on YelpChi, Amazon, and FDCompCN.","key_machinery":"The machinery is the heterophily separation module plus a dual-channel propagation step. The separator is a binary edge classifier that projects source and target features, computes $\\kappa_{uv} = \\tanh(W_h [h_u \\| h_v \\| (h_u - h_v)])$, and labels an edge heterophilic when $y_u \\neq y_v$, with a hinge-style auxiliary loss $L_H = \\frac{1}{|E_{tr}|} \\sum \\max(1 - \\kappa_{uv} y_{e_{uv}}, 0)$. The propagation step applies $W_f$ on the homophilic channel and $I - W_f$ on the heterophilic channel, so the two branches are complementary by construction; each branch adds a weighted residual of the original neighbor features and aggregates with degree-based re-scaling. The difference vector $z^+_u - z^-_u$ in the fusion layer is what lets the model learn to trust one channel over the other at each node.","core_discovery":"The central claim is that separating heterophilic from homophilic edges before message passing, rather than pruning or reweighting mixed neighborhoods, is what lets a GNN detect camouflaged fraudsters. DHMP trains an MLP edge classifier, guided by an auxiliary margin loss, to label each edge as same-label or cross-label; it then aggregates over the two resulting subgraphs in parallel using the complementary filters $W_f$ and $I - W_f$, so the two channels carry low-pass and high-pass components of the same signal. A re-scaled residual term preserves each neighbor's original features, degree normalization $\\sqrt{1 + d_u d_v}$ down-weights contributions from high-degree nodes, and a fusion layer concatenates $z^+$, $z^-$, and their difference so the model can weigh the two frequencies per node. The paper argues that because nodes adaptively balance the two channels and the whole pipeline is trained with balanced node and edge sampling, the learned embeddings separate fraudsters from benign users more cleanly, and it reports AUC gains over the strongest baselines of about 1.1 points on YelpChi, 0.4 on Amazon, and 3.0 on FDCompCN.","pith_inferences":["Beyond the paper: a natural stress test is to vary the train/test ratio of node labels and measure whether the separator's accuracy tracks final AUC; if it does, the method's gains are bottlenecked by edge-classification quality.","Beyond the paper: the complementary-filter construction $W_f$ versus $I - W_f$ suggests a spectral reading of DHMP as a learnable high/low-pass filter bank; one could test this by replacing $W_f$ with fixed graph filters and checking whether the gain disappears.","Beyond the paper: because the separator uses only node features and labels, not edge attributes, the approach may transfer to other heterophily problems such as protein interaction or citation networks where cross-label edges are informative; the paper does not claim this.","Beyond the paper: the method's reported success on FDCompCN, a small graph with only a few thousand nodes, hints that the framework may work where deep GNNs usually struggle; a test on larger and sparser graphs would reveal whether the gains scale."],"forward_implications":["If DHMP's claim holds, pruning or excluding heterophilic neighbors, as several baselines do, is not just unnecessary but harmful: the high-frequency channel carries information worth keeping.","Fraud graphs with multiple relation types each get separate homophilic and heterophilic splits, so the model can learn that one relation is more informative for spotting camouflage than another.","The auxiliary edge loss and balanced sampling give a recipe for training GNNs on heavily imbalanced node labels without oversampling the minority class.","The degree-based re-scaling should make the model more robust to spammers who create many edges, since high-degree neighbors are automatically down-weighted.","Because relation-level embeddings are concatenated after channel fusion, the framework can in principle be applied to any multi-relational graph where cross-label edges carry signal, not just the three datasets tested."],"supporting_citations":[{"why":"Supplies the FDCompCN financial-fraud graph and the spectral heterophily baseline SplitGNN that DHMP must beat.","marker":"[2]"},{"why":"Supplies the Amazon review-fraud dataset and H2-FDetector, the homophilic/heterophilic dual-propagation baseline DHMP most directly extends.","marker":"[3]"},{"why":"Supplies the YelpChi dataset and CARE-GNN, the neighbor-sampling baseline that prunes heterophilic edges.","marker":"[4]"},{"why":"Supplies FAGCN, the frequency-decomposition baseline that motivates separating low- and high-frequency signals.","marker":"[11]"},{"why":"Supplies BWGNN, the beta-wavelet baseline whose high-frequency filtering is the closest spectral competitor.","marker":"[13]"},{"why":"Supplies GFAN, the strongest AUC baseline on YelpChi and Amazon that DHMP reports beating.","marker":"[17]"},{"why":"Supplies SEFraud, the strongest Recall baseline on Amazon that DHMP reports beating.","marker":"[14]"}],"fun_headline_variants":["DHMP splits fraud graphs by edge type, beats 13 baselines","Dual-channel GNN separates same/cross-label edges for fraud detection","Graph fraud detection: heterophily-aware message passing lifts AUC","Splitting homophilic and heterophilic edges improves fraud AUC","DHMP: adaptive frequency fusion for graph fraud detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the edge-type classifier trained on node-label-derived edge labels generalizing to held-out edges, yet the paper reports no accuracy for this separator and no analysis of how its mistakes contaminate the two channels.","fun_headline_variants_meta":{"raw":{"variants":["DHMP splits fraud graphs by edge type, beats 13 baselines","Dual-channel GNN separates same/cross-label edges for fraud detection","Graph fraud detection: heterophily-aware message passing lifts AUC","Splitting homophilic and heterophilic edges improves fraud AUC","DHMP: adaptive frequency fusion for graph fraud detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1535,"prompt_tokens":1017,"completion_tokens":518,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":430}},"tokens_in":633,"tokens_out":518,"duration_ms":4603,"temperature":1.0,"reasoning_tokens":430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:53:46.258721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DHMP with the same settings but replace predicted edge types with ground-truth edge types on validation and test edges; if AUC does not rise, the separator is not the source of the gain. Conversely, shuffle edge-type labels during training while keeping the same loss; if performance holds, the dual-channel separation is not doing the claimed work. Either experiment would settle whether the mechanism is load-bearing.","supporting_citations":[{"cited_title":"H2-fdetector: A gnn-based fraud detector with homophilic and heterophilic connections,","cited_arxiv_id":null,"evidence_quote":"Supplies the Amazon review-fraud dataset and H2-FDetector, the homophilic/heterophilic dual-propagation baseline DHMP most directly extends."},{"cited_title":"Enhancing gnn-based fraud detector via semantic extraction and max-representation-margin,","cited_arxiv_id":null,"evidence_quote":"Supplies GFAN, the strongest AUC baseline on YelpChi and Amazon that DHMP reports beating."}],"review_version":1}