{"id":"ff1f289c-04b5-413b-aaec-811651b91e3a","arxiv_id":"2506.20980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RASH learns relation importance from a dual heterogeneous hypergraph, constructs homophilic and heterophilic views, and uses contrastive learning to improve heterogeneous graph representations.","lead":"This paper presents RASH, a self-supervised learning framework that separates homophilic and heterophilic patterns in heterogeneous graphs by learning relation importance from a dual hypergraph representation. The method reports consistent gains over existing graph learning approaches on four benchmark datasets across classification, clustering, and similarity search.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is the Eq. 9-10 product-of-importance heuristic, which is never validated against ground-truth labels; if low learned importance reflects noise rather than heterophily, the constructed views are arbitrary and the claimed separation may not be the source of the gains.","rationale":"The strongest claim is empirical: RASH's homophily-heterophily separation, driven by learned relation importance, improves node representations. For that claim to hold, the product of relation-importance weights along a two-hop path (Eq. 9) must correlate with same-label connectivity, and the product of complement weights (Eq. 10) must correlate with different-label connectivity. The paper provides no ground-truth validation of either correlation; it only shows end-task metrics and module ablations. If the relation-importance weights instead reflect edge reliability, frequency, or other confounds, the constructed homophilic and heterophilic graphs are arbitrary, and the contrastive objective may simply reinforce a self-generated structure. This is not an internal inconsistency but an unverified empirical premise, making it the most load-bearing assumption. I also considered the Eq. 8 inconsistency (delta sampled from Uniform(0,1) yet later set to 1e-4) and the self-referential positive-sample selection, but those are secondary: the label-correlation check is decisive for the central claim. Because the user is exactly the reader's weakest_assumption, and the reader's CONDITIONAL verdict already captures this unresolved risk, no verdict adjustment is needed.","tokens_in":19677,"tokens_out":5317,"duration_ms":59890,"concrete_test":"Using the released code on DBLP and IMDB, train RASH, freeze the model, and for each relation r compute a^{r,ho}_{i,j} and a^{r,he}_{i,j} (Eqs. 9-10) over same-type node pairs. Rank pairs by each weight and measure label agreement (for a^{ho}) and label disagreement (for a^{he}) among the top-k pairs used as neighbors, compared with random same-type pairs. If the top-k homophilic pairs are not significantly more label-agreeing than random, or the top-k heterophilic pairs are not significantly more label-disagreeing, the separation mechanism is unvalidated; then re-run the Table 4 ablations with random or degree-based replacement weights to test whether the gains actually depend on the learned relation importance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism is Section 4.3, Eqs. 9-10: a^{r,ho}_{i,j} = sum_k w^r_{i,k} w^{r-1}_{k,j} and a^{r,he}_{i,j} = sum_k (1-w^r_{i,k})(1-w^{r-1}_{k,j}) define which same-type node pairs are treated as homophilic or heterophilic. The whole claim that RASH 'adaptively separates homophilic and heterophilic patterns' rests on these weights actually tracking label agreement. The paper never validates this: there is no check that high-a^{ho} pairs connect same-label nodes more often than random pairs, or that high-a^{he} pairs connect different-label nodes. If learned low edge importance encodes noise or low relevance rather than heterophily, the homophilic and heterophilic graphs are self-generated artifacts, and the contrastive losses in Eqs. 13-16 will align representations to those artifacts. Reported gains could then be due to the heterogeneous encoder plus contrastive regularization rather than to genuine homophily-heterophily separation. This is load-bearing because the abstract and title attribute the performance exactly to that separation, and none of the ablations in Table 4 isolate the validity of the separation itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RASH, a self-supervised contrastive framework for heterogeneous graphs. It encodes relation-specific bipartite subgraphs through a dual heterogeneous hypergraph, derives edge importance weights via a Gumbel-style soft sampling (Eq. 8), and constructs homophilic and heterophilic graphs by multiplying learned importance scores along two-hop paths (Eqs. 9-10). Low- and high-pass filters extract view representations, and a multi-relation InfoNCE loss (Eqs. 13-16) aligns these views with the heterogeneous encoder output. Experiments on DBLP, ACM, IMDB, YELP, and a large-scale Aminer graph report state-of-the-art or competitive results in node classification, clustering, and similarity search, along with ablations, robustness to edge deletion, randomized-feature tests, and hyperparameter sensitivity.","tokens_in":19963,"tokens_out":7329,"duration_ms":79427,"significance":"If the central assumption of Eqs. 9-10 is valid, RASH is a meaningful step: it addresses heterophily in heterogeneous graphs without collapsing the graph to predefined meta-paths, and it does so in an end-to-end, self-supervised manner. The paper's strengths include public code, consistent gains across four benchmark datasets plus a 439k-node large-scale experiment, a sensible ablation suite, and robustness analyses. The significance is currently conditional, however, because the paper never tests whether the constructed homophilic and heterophilic graphs correspond to ground-truth label agreement. That validation is the difference between a principled separation mechanism and a self-generated artifact; the authors should be required to provide it.","major_comments":[{"comment":"The central assumption that a^{r,ho}_{i,j} and a^{r,he}_{i,j} track ground-truth label agreement among target nodes is never validated. A direct test would rank same-type node pairs by these scores and measure whether high-ho pairs connect same-label nodes and high-he pairs connect different-label nodes more often than chance, for example by reporting label-agreement precision or AUC. The ablations in Table 4 remove whole modules (w/o Homo_CL, w/o Hete_CL, w/o RAE) but do not isolate whether the constructed graphs themselves are meaningful. If low learned importance encodes noise or irrelevance rather than heterophily, the contrastive losses in Eqs. (13)-(16) may align representations to self-generated artifacts, and the reported gains could stem from the heterogeneous encoder plus contrastive regularization rather than from genuine homophily/heterophily separation. This is load-bearing because the title and abstract attribute the performance gains to exactly this separation.","section":"Section 4.3, Eqs. (9)-(10)"},{"comment":"There is an internal contradiction in the treatment of the noise parameter δ. Eq. (8) defines δ ~ Uniform(0,1) and uses log δ - log(1-δ), which is a logistic (Binary Concrete) noise term, not a Gumbel variate as claimed; Section 5.3 then states that 'δ was set to 1e-4', which makes the noise a constant and removes the stochasticity that the reparameterization claim requires. Please correct the formula or the hyperparameter description, and state explicitly whether w is stochastic or deterministic. This matters for reproducibility and for the claim that the model learns discrete edge sampling via Gumbel-Max reparameterization.","section":"Section 4.2.2, Eq. (8) and Section 5.3"}],"minor_comments":[{"comment":"Equation (3) contains typos in the incidence-matrix definition: the relation index should appear on the entries (M^r_{i,e}, M^r_{j,e}), the condition 'iff A^r_{i,j}=1, <i,j>=e' is malformed, and the following sentence 'M^r denote the transpose matrix' should read 'M^{r\\top}'.","section":"Section 4.2.1, Eq. (3)"},{"comment":"Equation (9) sums over k ∈ N^r_j, but the described two-hop path i → k under relation r and k → j under r^{-1} requires k to be in the intersection of the relevant neighbor sets; please clarify the intended index sets for both Eq. (9) and Eq. (10).","section":"Section 4.3, Eq. (9)"},{"comment":"The sentence 'The range of the number of positive samples was set from 0 to 5' appears inconsistent with the top-k positive sampling described in Section 4.4; with k=0 no positive sample would exist, so the searched range is presumably 1 to 5.","section":"Section 5.3"},{"comment":"The text says features are drawn from the 'Xavier uniform distribution [37]', but reference [37] is the heterogeneous graph convolution paper, not the original source of Xavier initialization; please fix the citation.","section":"Section 5.7 and Reference [37]"},{"comment":"The axis label 'Value of c' in Figure 4(b) should read τ_c, and the labels 'llow' and 'lhigh' in Figure 5 are undefined; please provide explicit axis labels and a legend or caption definition.","section":"Figures 4 and 5"},{"comment":"The statement that RASH's NMI and ARI on YELP are 'improved by 13.74% and 14.21%' over the second-best method should be phrased as percentage-point differences (i.e., 74.93 vs. 61.19 and 77.08 vs. 62.87), to avoid confusion with relative improvements.","section":"Section 5.5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eqs. (9)-(10) is well-founded and should be the primary request in the revision. I am not recommending rejection because the issue is addressable within the paper's scope: the authors can add a label-agreement validation of the constructed homophilic and heterophilic graphs. If that validation fails, however, the central claim collapses and the paper should not be accepted. The Gumbel/δ inconsistency is a separate correctness issue that must also be fixed. The paper fits the scope of the venue and has clear strengths in experimental breadth and code availability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RASH is a plausible, well-engineered method with a genuinely new mechanism: it learns relation importance via a dual heterogeneous hypergraph and uses it to build homophilic and heterophilic views for contrastive learning. The experimental results are consistent across four benchmarks and several tasks, with public code and a sensible random-feature ablation. That is more than many papers in this area deliver.\n\nThe soft spots are real but mostly addressable. The central assumption in Eq 9-10—that the product of learned edge importances along a two-hop path tracks label agreement, and the product of complements tracks label disagreement—is never checked against ground truth. The paper never shows that high-weight homophilic pairs are actually more likely to share labels. If low importance encodes noise or low relevance rather than heterophily, the constructed views are artifacts, and the contrastive loss could be aligning to those artifacts. The ablations show that removing either loss hurts, but that does not establish that the separation is correct.\n\nThere is also an internal inconsistency: Eq 8 describes delta as sampled from Uniform(0,1), while Section 5.3 fixes delta to 1e-4. That is probably a typo, but it matters for how to read the Gumbel reparameterization claim. Separately, Hetero2Net and HDHGR are discussed in related work but not included in the benchmarks; they are the closest prior art and their absence weakens the comparison.\n\nOn the citation side, the paper is reasonably honest about prior work; the 'first time' claim in the introduction is a bit strong, but the mechanism is new enough that it doesn't bother me much.\n\nWho is this for? People working on heterogeneous graph self-supervised learning and heterophily. It is worth a serious referee, but the review should push for a direct validation experiment: compute the constructed homophilic/heterophilic edge sets and compare their label-agreement rates against random pairs. That one experiment would settle the main concern.","headline":"RASH is a genuinely new mechanism with strong benchmark results, but its central homophily/heterophily separation is an unvalidated heuristic that needs a direct label-agreement check before the claim is credible.","tokens_in":20503,"tokens_out":2382,"would_cite":true,"duration_ms":26882,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Heterogeneous graphs can separate homophily from heterophily by learning relation importance, and this paper claims the resulting representations beat existing methods on node classification, clustering, and similarity search.","keywords":["Heterogeneous graph neural networks","Heterophily","Self-supervised learning","Contrastive learning","Dual heterogeneous hypergraph","Relation importance","Node classification","Graph representation learning"],"falsifier":"Train RASH on one benchmark, extract the node pairs ranked most homophilic by equation (9) and most heterophilic by equation (10), and compare the fraction of each set that actually shares ground-truth node labels; if the 'homophilic' set is not clearly more label-agreeing than the 'heterophilic' set—or no better than random edge weights—the separation mechanism is not learning homophily.","tokens_in":19426,"feed_emoji":"🕸️","tokens_out":9603,"duration_ms":90758,"temperature":0.7,"pith_summary":"Heterogeneous graphs—networks with multiple node and edge types—often connect nodes that carry different labels, a property known as heterophily. Existing graph-learning methods usually flatten such graphs into homogeneous ones before trying to model heterophily, and in doing so lose the meaning carried by the different relation types. This paper tries to establish that both heterogeneity and heterophily can be handled together: a contrastive framework called RASH learns how important each heterogeneous relation is, uses that importance to split same-type node pairs into homophilic (same-label) and heterophilic (different-label) groups, and aligns the resulting views with the original graph representations. If right, it would mean that predefined meta-paths and homogeneous simplification are not needed to capture heterophily, and that one learned set of relation weights can serve multiple node types and downstream tasks. On DBLP, ACM, IMDB, and YELP, RASH reports the best node-classification, clustering, and similarity-search numbers among the compared methods.","feed_headline":"Relation importance separates homophily from heterophily at edge level","feed_subtitle":"Edge weights split homophily from heterophily, improving node classification and clustering on four benchmarks.","key_machinery":"The load-bearing object is the dual heterogeneous hypergraph transform (DHHT) together with relation-aware importance scoring. In DHHT, each relation's bipartite subgraph is written as an incidence matrix whose columns are original edges, so original edges become hyperedges; hypergraph convolution then learns edge-level representations via two-stage message passing between nodes and hyperedges, capturing higher-order interaction among heterogeneous edges. A scoring function converts those representations into edge importance weights through Gumbel-Softmax sampling. Homophilic and heterophilic graphs are constructed dynamically from two-hop products of the importance weights and their complements, and a multi-relation contrastive loss aligns the low-pass filtered homophilic view, the high-pass filtered heterophilic view, and the original heterogeneous view in a shared space.","core_discovery":"The central claim is that heterophily in a heterogeneous graph can be learned from the graph's own relational structure rather than from a homogeneous projection of it. RASH converts each relation-specific bipartite subgraph into a dual heterogeneous hypergraph, runs hypergraph convolution to obtain edge representations, and assigns each edge a Gumbel-softmax importance weight. For a pair of same-type nodes, the two-hop product of these weights (equation 9) is treated as a homophilic edge weight, and the product of the complement weights (equation 10) as a heterophilic edge weight. Low-pass filtering on the homophilic graph and high-pass filtering on the heterophilic graph produce views that are pulled toward the heterogeneous encoder by a multi-relation InfoNCE-style contrastive loss. The paper reports that this yields state-of-the-art node classification, clustering, and similarity search on DBLP, ACM, IMDB, and YELP, and that the gains persist when node features are replaced by random vectors.","pith_inferences":["An untested corollary of the two-hop construction is that equation (9) could serve as a general-purpose heterophily detector for heterogeneous graphs, usable even when no labels or downstream task is specified.","A direct test of the mechanism would correlate learned edge importance with ground-truth label agreement; if the correlation is weak, the contrastive loss may be aligning with graph structure rather than with homophily.","The dual-hypergraph edge encoder is detachable and could be inserted into supervised heterogeneous graph models, not only contrastive pipelines.","If low importance weights mostly mark noisy or irrelevant relations rather than heterophily, equation (10) would need an explicit irrelevance term; this is the part of the design a follow-up should examine."],"forward_implications":["Because relation importance is learned end-to-end, no predefined meta-paths are needed to separate homophilic from heterophilic structure.","One set of learned importance weights yields homophilic and heterophilic views for any target node type without retraining the model for each type.","The dual hypergraph supplies structural signal even when node features are replaced with random vectors, so the method is not dependent on informative raw features.","Under random edge deletion, the reported classification accuracy degrades more slowly than the compared methods, indicating that the separated views provide redundancy against missing edges."],"supporting_citations":[{"why":"Supplies the hypergraph convolution used to propagate messages between nodes and hyperedges in the dual heterogeneous hypergraph.","marker":"[5]"},{"why":"Defines the dual-hypergraph and incidence-matrix construction that turns original edges into hyperedges.","marker":"[14]"},{"why":"Provides the Gumbel-Softmax reparameterization that makes edge-importance sampling differentiable.","marker":"[11]"},{"why":"Supplies the simplified graph-convolution encoding used for low-pass and high-pass filtering on the constructed graphs.","marker":"[32]"},{"why":"Provides the InfoNCE contrastive objective that aligns the heterogeneous, homophilic, and heterophilic views.","marker":"[39]"},{"why":"Is the heterophily-aware latent-graph baseline whose preprocessing step RASH replaces with end-to-end learned relation importance.","marker":"[26]"},{"why":"Is the heterophily baseline that assumes metapath-connected nodes are homophilic, the assumption RASH argues is too restrictive.","marker":"[15]"},{"why":"Is the metapath-based heterophily rewiring method contrasted when RASH argues for keeping heterogeneous relations intact.","marker":"[8]"}],"fun_headline_variants":["Relation-aware contrastive learning separates homophily from heterophily","RASH: Dual hypergraphs split homophily and heterophily in heterogeneous graphs","Learning heterophily from relations, not homogeneous projections","Contrastive hypergraphs separate homophilic and heterophilic edges","Edge importance weights separate homophily and heterophily"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that along a two-hop path, the product of learned edge-importance weights marks node pairs that share a label, while the product of the complement weights marks node pairs with different labels; this label-interpretation is assumed, not verified against ground-truth edge labels.","fun_headline_variants_meta":{"raw":{"variants":["Relation-aware contrastive learning separates homophily from heterophily","RASH: Dual hypergraphs split homophily and heterophily in heterogeneous graphs","Learning heterophily from relations, not homogeneous projections","Contrastive hypergraphs separate homophilic and heterophilic edges","Edge importance weights separate homophily and heterophily"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000426,"raw_usage":{"total_tokens":2210,"prompt_tokens":999,"completion_tokens":1211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1122}},"tokens_in":615,"tokens_out":1211,"duration_ms":9068,"temperature":1.0,"reasoning_tokens":1122,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:37:14.331397+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train RASH on one benchmark, extract the node pairs ranked most homophilic by equation (9) and most heterophilic by equation (10), and compare the fraction of each set that actually shares ground-truth node labels; if the 'homophilic' set is not clearly more label-agreeing than the 'heterophilic' set—or no better than random edge weights—the separation mechanism is not learning homophily.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the dual-hypergraph and incidence-matrix construction that turns original edges into hyperedges."},{"cited_title":"When Heterophily Meets Heterogeneous Graphs: Latent Graphs Guided Unsupervised Representation Learning","cited_arxiv_id":"2409.00687","evidence_quote":"Is the heterophily-aware latent-graph baseline whose preprocessing step RASH replaces with end-to-end learned relation importance."},{"cited_title":"Hetero$^2$Net: Heterophily-aware Representation Learning on Heterogenerous Graphs","cited_arxiv_id":"2310.11664","evidence_quote":"Is the heterophily baseline that assumes metapath-connected nodes are homophilic, the assumption RASH argues is too restrictive."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the metapath-based heterophily rewiring method contrasted when RASH argues for keeping heterogeneous relations intact."}],"review_version":1}