{"id":"53c3493d-f7aa-4624-b31b-8796013a1f00","arxiv_id":"2506.06682","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"HetCRF combines masked autoencoding and contrastive learning in a dual-channel framework with two positive-sample augmentation strategies, improving heterogeneous graph node classification under sparse features.","lead":"This paper introduces HetCRF, a self-supervised method for heterogeneous graphs that trains a shared encoder with both masked reconstruction and contrastive learning. The authors report higher node classification accuracy on four benchmark graphs, especially when node features are missing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (26) defines the contrastive loss with self-similarities in the denominator, so the specified objective is not a valid contrastive loss and the gradient-balance theorem does not apply to it.","rationale":"The central claim is that HetCRF's contrastive-reconstruction design gives state-of-the-art results. For this to be true, the model must be specified by equations that are actually trained. Eq. (26), the only definition of L_con, uses self-similarities in the denominator; this is not a valid contrastive loss and has different gradients. The paper's Theorem 1 and its generalization in Eq. (21) are also wrong for multiple positives, which directly undermines the stated motivation for the two positive-augmentation strategies. This is an internal inconsistency, not a disagreement with consensus. It is load-bearing because the authors' explanation of why their method works (gradient balancing) rests on it, and because the reported numbers cannot be reproduced from the text. I considered the reader's clustering-assumption concern, which is plausible, but the Eq. (26)/Eq. (21) issue is more basic: even with clean clusters, the loss as written is not contrastive. I also note the paper's own Table 1 shows ASHGCL and HGMS-C beating HetCRF on DBLP Ma-F1@20 and ACM Ma-F1@40, so the broad 'outperforms SOTA' claim is overstated; however the specific 40%-label feature-missing claims may survive a corrected implementation. No code or hyperparameters are given, so the concrete test is the only way to verify the reported results.","tokens_in":20950,"tokens_out":6517,"duration_ms":68959,"concrete_test":"Re-implement the contrastive channel exactly as Eq. (26) is printed and run the reported AMiner 40%-label node-classification protocol, then rerun with the corrected denominator Σ_{j∈P_i∪N_i} exp(Z_i^fusion · Z_j^schema/τ). If the two Macro-F1 values differ by more than the reported standard deviation, the paper's stated objective is not the objective that produced its headline results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The only contrastive objective, Eq. (26), is written as -log[ Σ_{k∈P_i} exp(Z_i^fusion · Z_k^schema/τ) / Σ_{j∈P_i∪N_i} exp(Z_j^fusion · Z_j^schema/τ) ]. The denominator uses a node's own embedding in both factors (Z_j·Z_j), not the anchor-to-sample similarity Z_i·Z_j. This is not a valid softmax over similarities to the anchor; it is a function of per-node norms, so the reported gradient-balance theorem does not describe the objective being optimized. In addition, the claimed generalization in Eq. (21) is false for multiple positives: for m positives, Σ_{p∈P} ∂L/∂f_p = -(m - Σ_{p∈P} p_p) f_i/τ while Σ_{n∈N} ∂L/∂f_n = (1 - Σ_{p∈P} p_p) f_i/τ, equal only when m=1. Since both positive-augmentation strategies are justified by that equality, the theoretical motivation for the paper's key novelty is unsound as written. Either the equation is a typo that must be corrected, or the experiments were run with a different objective than the one specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HetCRF, a self-supervised dual-channel framework for heterogeneous graphs. The generative channel performs masked feature reconstruction and meta-path reconstruction with a shared HAN encoder; the contrastive channel builds views from encoder embeddings (schema view and fusion view) and passes them through a GCN for secondary aggregation. Two positive-sample augmentation strategies are introduced, one based on meta-path connection counts and one based on K-means clustering of encoder embeddings, with the stated goal of balancing positive and negative gradient contributions in the contrastive loss. Experiments on DBLP, Freebase, ACM, and AMiner report gains in node classification and clustering, with the largest claimed gains on feature-missing datasets such as AMiner and Freebase.","tokens_in":21267,"tokens_out":8891,"duration_ms":83723,"significance":"If the claims held, HetCRF would be a useful contribution to heterogeneous-graph self-supervision, particularly in sparse-feature regimes: the paper targets a real limitation of existing hybrid frameworks and provides ablations on four datasets. However, the theoretical gradient-balance argument is not correctly connected to the contrastive loss as written in Eq. (26), and the ablation results in Table 3 do not consistently support the claim that the two augmentation strategies are jointly beneficial. These are load-bearing issues for the paper's central contribution, so the current version cannot be accepted. I also note that no code or full hyperparameter settings are provided, which limits independent verification.","major_comments":[{"comment":"The contrastive loss in Eq. (26) is written with a denominator Σ_{j∈P_i∪N_i} exp(Z_j^fusion · Z_j^schema / τ), i.e., self-similarities rather than anchor-to-sample similarities. This is not a valid contrastive softmax over the anchor's neighbors, and it does not reduce to the InfoNCE form used in Theorem 1. Consequently, the gradient-balance theorem and its generalization in Eq. (21) do not apply to the objective as specified. Please correct the denominator to Σ_{j∈P_i∪N_i} exp(Z_i^fusion · Z_j^schema / τ), or state explicitly that a different objective was used in the experiments. This is not a cosmetic issue: the theoretical motivation for both positive-augmentation strategies depends on it.","section":"§3.2.2, Eq. (26)"},{"comment":"For a genuine multi-positive InfoNCE loss L_i = -log[Σ_{p∈P} exp(s_ip/τ) / (Σ_{p∈P} exp(s_ip/τ) + Σ_{n∈N} exp(s_in/τ))], the identity Σ_{p∈P} ∂L_i/∂f_p = -Σ_{n∈N} ∂L_i/∂f_n is actually correct, so I do not dispute the generalization itself. However, the paper does not show the derivation, and the identity fails for the loss as written in Eq. (26). Please provide the multi-positive derivation explicitly and state it for the corrected loss; the current 'analogous proof' is insufficient because the denominator terms in Eq. (26) change the relevant derivatives.","section":"§3.2.2, Eq. (21)"},{"comment":"Table 3 does not support the claim that the two augmentation strategies are jointly beneficial. On DBLP, w/o_Cluster (MPC-only) achieves 92.75 Macro-F1 versus 92.55 for the full HetCRF; on ACM, w/o_MPC (cluster-only) achieves 92.51 versus 92.35 for the full model. Thus, adding the second strategy degrades performance on two of the four datasets. The paper should either explain this inconsistency or revise the claim that both strategies contribute to the final model.","section":"§4.3.2, Table 3"},{"comment":"The clustering-based augmentation uses the true number of classes S in K-means during self-supervised pre-training. If S is derived from label information, the method is not label-free in the usual self-supervised sense, and the comparison to baselines is unfair unless all baselines also use S. Moreover, noisy cluster assignments can inject false positives; the paper provides no experiment varying S or measuring cluster quality. Please justify the use of S or show robustness to misspecified S.","section":"§3.2.2, Positive Sample Augmentation via Clustering Algorithm"},{"comment":"The statement that HetCRF demonstrates superior performance over the existing baselines in clustering is not accurate for all datasets and metrics. On ACM, HGMS-C achieves NMI 71.97 while HetCRF achieves 69.48, so HetCRF is not the best on that dataset. The clustering claim should be qualified to reflect the actual wins and losses in Table 2.","section":"§4.2, Table 2"}],"minor_comments":[{"comment":"The text says the three training strategies were evaluated under a 20% label rate, while Figure 3's caption says 'Macro-F1, 40%'; please make the label rate consistent between text and figure.","section":"§4.3.1"},{"comment":"Notation is inconsistent: Eq. (5) uses σ for the activation, while Eq. (7) uses δ; Eq. (9) has |Z_i| in the softmax denominator, which should presumably be |Φ|; and the text near Eq. (5) says 'Hermitian space', which appears to mean the node-type feature space.","section":"§3.1 and §3.2.1"},{"comment":"Figures 3 and 4 are difficult to read: the t-SNE labels in Figure 4 are illegible, and Figure 3 renders axis labels as encoded tokens rather than readable text. Please provide vector graphics with readable fonts.","section":"Figures 3 and 4"},{"comment":"No code, hyperparameter table, or training details are provided, so the numerical results in Tables 1-4 cannot be independently reproduced. Please include a reproducibility appendix with exact settings and random-seed handling.","section":"§4"},{"comment":"The reference list says 'HERO ICRL'24', which should be 'ICLR'24'; also, the baseline enumeration in Section 4 omits DGI and DMGI, which appear in Table 2, so the list of compared methods in the text does not match the tables.","section":"References and baselines"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the load-bearing issue is Eq. (26). If that equation reflects the implemented objective, the main theoretical contribution is not valid; if it is a typo, the authors must correct it and re-derive the gradient-balance claims. I would also like to see the Table 3 inconsistency addressed before any acceptance decision, and I recommend asking for the code or a detailed reproducibility appendix, since several baseline numbers are quoted from other papers without independent runs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HetCRF is worth a look for anyone working on hybrid generative-contrastive SSL for heterogeneous graphs. The two-view construction from encoder embeddings is a sensible response to feature-sparse settings, and the reported gains, while modest, are consistent across four datasets. But the paper is not in publishable shape: the contrastive loss in Eq. (26) has self-similarities in the denominator, so the objective is not a valid softmax over anchor-to-sample similarities, and the gradient-balance theorem the authors prove does not apply to it. That is load-bearing, because the positive-augmentation strategies are justified by that theorem.\n\nWhat the paper does well: it identifies a real mismatch between generative and contrastive channels sharing an encoder, and it proposes a concrete fix by adding a GCN for secondary aggregation and constructing views from encoder embeddings rather than raw features. The schema view (structural attention over heterogeneous neighbors) and fusion view (semantic attention over meta-paths) are clearly described. The two positive-augmentation heuristics are ablated, and the ablation showing the GCN view construction helps most on missing-feature datasets (AMiner, Freebase) supports the core claim. The authors also honestly state that GC-HGNN was not compared because its code is unavailable; that is a limitation, not a concealment.\n\nThe main flaw is Eq. (26). The denominator sums exp(Z_j^fusion · Z_j^schema/τ) over all j in P_i ∪ N_i; it never involves the anchor i, so it is a constant normalizer. The loss neither pushes negatives away nor matches the InfoNCE form the theorem assumes. This might be a typo—the standard denominator would be Σ_j exp(Z_i^fusion · Z_j^schema/τ)—but as written it breaks the theoretical motivation. One stress-test claim you will hear, that Eq. (21) is false for multiple positives, is wrong: a careful derivation shows the sum of positive gradients equals the sum of negative gradients in magnitude for any number of positives. So the theory, once the loss is corrected, is fine. Separate issues: no hyperparameters or training details are given, no code is released, and the clustering augmentation requires knowing the true class count S. These are fixable in revision.\n\nWho should read this: people thinking about shared encoders or about contrastive views in feature-scarce HINs. I would send it to peer review, not desk reject, but with a clear request for major revision: fix the loss, report the missing experimental details, and if possible add the GC-HGNN comparison.","headline":"HetCRF is a promising hybrid SSL architecture for heterogeneous graphs that is not review-ready because the stated contrastive loss is internally inconsistent, though the issue looks like a fixable typo.","tokens_in":21772,"tokens_out":7751,"would_cite":false,"duration_ms":74708,"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":"The paper claims that a single HAN encoder shared between masked reconstruction and contrastive learning, with contrastive views built from embeddings plus two positive-sample augmentation strategies, outperforms state-of-the-art…","keywords":["self-supervised learning","contrastive learning","masked autoencoder","heterogeneous graph","graph neural network","positive sample augmentation","meta-path","semantic sparsity"],"falsifier":"On AMiner at a 40% label rate, replace the K-means positive augmentation with randomly sampled positives of the same set size and keep everything else fixed; if Macro-F1 stays near 79.68, the semantic validity of the clusters is not doing the work, whereas a clear drop would confirm the assumption. A second observation: run the same experiment with the cluster count set to a value different from the known class count; if performance is unchanged, the method does not actually depend on knowing $S$.","tokens_in":20733,"feed_emoji":"🕸️","tokens_out":6360,"duration_ms":53745,"temperature":0.7,"pith_summary":"The paper claims that on heterogeneous graphs, a single shared encoder can serve both generative reconstruction and contrastive learning, provided the contrastive branch is rebuilt on top of encoder embeddings instead of raw inputs. Its proposed framework, HetCRF, pairs masked feature and meta-path reconstruction with a GCN-augmented contrastive channel, and balances the contrastive loss by augmenting positive samples through meta-path connectivity and K-means clustering. The payoff would be node representations that survive semantically sparse settings, where raw features are missing and ordinary view augmentation destroys what little signal exists. On AMiner and Freebase at a 40% label rate, the paper reports Macro-F1 gains of 2.75 and 2.2 points over the second-best baseline.","feed_headline":"One shared encoder beats sparse-graph baselines on missing features","feed_subtitle":"HetCRF builds contrastive views from embeddings, not raw features, and gains 2.75 Macro-F1 on AMiner at 40% labels.","key_machinery":"The framework's load-bearing object is the shared HAN encoder, a heterogeneous graph attention network that maps multiple node and edge types into common embeddings, reused by both channels. The generative channel masks node features and meta-path edges and reconstructs them under scaled cosine error; the contrastive channel takes the encoder's embeddings, applies PathSim-based top-K neighbor selection and GCN propagation to form schema and fusion views, and optimizes a contrastive loss with a learned positive-sample set. That set is built by counting connecting meta-paths across k hops and by K-means on the encoder's own embeddings, selecting 'key deviated nodes' per cluster as extra positives; the final loss is a weighted sum of the two reconstruction losses and the contrastive loss.","core_discovery":"The central claim is that the apparent incompatibility between masked autoencoding and contrastive learning under a shared encoder disappears if the contrastive channel performs a second aggregation step on the encoder's output. HetCRF builds two contrastive views, a schema view from heterogeneous-neighbor aggregation and a fusion view from attention-weighted meta-path adjacency, both propagated through a GCN, so contrastive semantics are drawn from embeddings rather than from brittle raw-feature augmentations. In addition, the paper proves that in InfoNCE the positive sample's gradient equals the sum over all negatives, and argues that adding extra positives rebalances the gradient and protects global-structure learning. The experimental claim is that this design outperforms state-of-the-art generative, contrastive, and hybrid baselines across DBLP, ACM, AMiner, and Freebase, with its largest margins precisely on the two feature-missing datasets.","pith_inferences":["A random-positive control of equal cardinality would separate the gradient-balancing effect from the semantic quality of K-means positives; the paper's ablations do not include one, so the reported gains could partly come from having more positives rather than from the specific chosen positives.","Because the clustering augmentation needs the true number of classes $S$ at pre-training time, applying HetCRF to a graph with an unknown class count is a risk the paper does not address; varying $S$ in the AMiner setting is a cheap test of how load-bearing this assumption is.","The gradient-balance theorem is stated for the standard InfoNCE objective, so the same positive-augmentation logic could be lifted into any InfoNCE-based graph contrastive learner without the rest of HetCRF.","The two attention-weighted view constructions give a template for handling sparse semantics; one could try building the same schema/fusion views on top of a homogeneous masked-autoencoder encoder, transferring the mechanism beyond heterogeneous graphs."],"forward_implications":["On feature-missing heterogeneous graphs such as AMiner and Freebase, the reported gains (Macro-F1 +2.75 and +2.2 points at 40% labels) are larger than on feature-rich graphs, so embedding-level view construction appears to be the regime where the design pays off most.","The ablation shows that the hybrid beats generative-only and contrastive-only training, and removing the GCN/view-construction step costs up to 5.04 and 4.05 Macro-F1 points on the two sparse datasets.","Positive augmentation itself matters: with neither augmentation strategy, Macro-F1 drops by several points at 40% labels on Freebase and AMiner, and the two strategies combined give the best numbers.","When original node features are removed from DBLP and ACM, HetCRF still leads four strong baselines, which is the paper's evidence that the framework generalizes to semantically sparse conditions.","The gradient-balance theorem implies that any InfoNCE-based contrastive channel can rebalance positive and negative contributions by enlarging the positive set, not just this particular architecture."],"supporting_citations":[{"why":"Supplies the masked-feature reconstruction design and scaled cosine error used by HetCRF's generative channel.","marker":"[8]"},{"why":"Defines the InfoNCE loss whose positive-versus-negative gradient imbalance Theorem 1 analyzes.","marker":"[19]"},{"why":"Provides the PathSim top-K neighbor similarity used to construct the schema and fusion views.","marker":"[24]"},{"why":"Introduces meta-path masking and feature reconstruction for heterogeneous graphs, the generative baseline HetCRF extends and compares against.","marker":"[26]"},{"why":"Defines the HAN encoder that HetCRF shares between its generative and contrastive channels.","marker":"[31]"},{"why":"The co-contrastive heterogeneous graph baseline that anchors the contrastive comparison.","marker":"[32]"},{"why":"GCMAE, the shared-encoder hybrid framework whose limitations motivate the GCN-augmented second aggregation in HetCRF.","marker":"[34]"},{"why":"Supplies the K-means clustering used in the clustering-based positive-sample augmentation strategy.","marker":"[17]"}],"fun_headline_variants":["HetCRF merges MAE and contrastive learning for heterogeneous graphs","Dual-channel framework beats baselines on sparse feature graphs","Embedding-based views fix contrastive learning under sparse semantics","Second aggregation step makes shared encoder work for MAE and CL","HetCRF gains 2.75 Macro-F1 on AMiner with only 40% labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the clustering augmentation, the model assumes it knows how many true classes $S$ exist and that K-means on its own current embeddings produces semantically valid positive pairs, so noisy clusters would inject false supervision into contrastive training.","fun_headline_variants_meta":{"raw":{"variants":["HetCRF merges MAE and contrastive learning for heterogeneous graphs","Dual-channel framework beats baselines on sparse feature graphs","Embedding-based views fix contrastive learning under sparse semantics","Second aggregation step makes shared encoder work for MAE and CL","HetCRF gains 2.75 Macro-F1 on AMiner with only 40% labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1339,"prompt_tokens":983,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":599,"tokens_out":356,"duration_ms":3799,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:52:16.326063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On AMiner at a 40% label rate, replace the K-means positive augmentation with randomly sampled positives of the same set size and keep everything else fixed; if Macro-F1 stays near 79.68, the semantic validity of the clusters is not doing the work, whereas a clear drop would confirm the assumption. A second observation: run the same experiment with the cluster count set to a value different from the known class count; if performance is unchanged, the method does not actually depend on knowing $S$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PathSim top-K neighbor similarity used to construct the schema and fusion views."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GCMAE, the shared-encoder hybrid framework whose limitations motivate the GCN-augmented second aggregation in HetCRF."}],"review_version":1}