{"id":"211227be-a99d-434a-b71c-8e7c8251dfd9","arxiv_id":"2412.15822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"S2DN improves inductive knowledge graph completion by smoothing similar relations and pruning unreliable edges, reporting gains over several baselines on WN18RR, FB15k-237, and NELL-995.","lead":"This paper introduces S2DN, a graph neural network method for completing knowledge graphs when new entities appear, by blurring relations with similar meaning and filtering out noisy edges. The authors report higher accuracy than several existing methods on standard benchmarks, though strong baselines are omitted.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'surpasses SOTA' claim is not established: strong baselines (AdaProp, RED-GNN, GraIL) are omitted, and the paper's own tables show S2DN behind RMPI on NELL-995 V4 and WN18RR V4.","rationale":"The reader's stated weakest assumption concerns the structure refining module's edge-reliability estimate. That is a legitimate technical risk, but the ablations provide some support for the module, and it is not the first thing that threatens the paper's central claim. The paper's headline claim is empirical: S2DN surpasses state-of-the-art models. That claim fails on the evidence as reported. The main tables omit the strongest current baselines in the inductive KGC literature, and the evaluation protocol is explicitly different from the protocol used by AdaProp and RED-GNN, making the reported numbers non-comparable to published SOTA results. The paper's own NELL-995 V4 and WN18RR V4 rows already show RMPI beating S2DN on key metrics. If a full-ranking comparison with AdaProp and RED-GNN is added and S2DN still wins, the concern is resolved and the conditional accept can stand. If not, the abstract and conclusion must be revised to a more limited claim. This does not require changing the reader's conditional verdict; it sharpens the reason for it.","tokens_in":22437,"tokens_out":4811,"duration_ms":43463,"concrete_test":"Run S2DN with the released code under AdaProp/RED-GNN's full filtered ranking protocol on WN18RR, FB15k-237, and NELL-995 V1–V4, and add published or reimplemented GraIL, RED-GNN, and AdaProp to the main comparison tables. If S2DN does not beat RED-GNN and AdaProp in Hits@10 and MRR, the SOTA claim must be retracted. Separately, rerun NELL-995 V4 with official hyperparameters to check whether the reported 74.89 versus RMPI's 84.20 gap reproduces.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that S2DN 'surpasses the performance of state-of-the-art models' (Abstract; Conclusion)—is contradicted by the paper's own results. In Table 9 (NELL-995 V4), S2DN achieves Hits@10 74.89 and MRR 64.12, while RMPI achieves 84.20 and 72.63; in Table 1 (WN18RR V4), RMPI's Hits@10 82.33 exceeds S2DN's 82.15. More fundamentally, the comparison omits the strongest inductive KGC baselines. Appendix B.2.3 excludes RED-GNN and AdaProp solely because they use a different evaluation protocol (ranking over all entities rather than 50 sampled negatives, Appendix B.1.1), and GraIL—the foundational subgraph method—is absent from the main tables. Because the reported numbers are generated under the 50-negative sampling protocol, they are not directly comparable to the published SOTA results, so 'surpasses SOTA' is unsupported as stated. This is a correctness risk for the headline claim, not merely a disagreement with community consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes S2DN, a subgraph-based inductive knowledge graph completion model that combines two modules: a semantic smoothing module that blurs relation embeddings with similar semantics through a Gumbel-softmax categorization, and a structure refining module that estimates edge reliability as a Bernoulli probability from MLP-transformed node features and prunes low-weight edges. The smoothed subgraph is processed by an RGNN and the refined subgraph by a GCN; the two global readouts are concatenated and scored by a classifier. Experiments on WN18RR, FB15k-237, and NELL-995, each with four inductive splits, report Hits@1, Hits@10, and MRR, together with ablations and robustness tests under injected semantic and structural noise. The paper claims that S2DN surpasses state-of-the-art models.","tokens_in":22756,"tokens_out":6718,"duration_ms":56974,"significance":"If the reported gains are reproducible, the two modules are a plausible and generally applicable contribution to inductive KGC; the released code, the use of public benchmarks, the ablations of both modules, and the five-run repetition are strengths, and I did not find evidence of circular fitting to the test labels. However, the headline 'surpasses SOTA' claim is not supported by the paper's own tables, and the baseline comparison is incomplete and uses a protocol that prevents direct comparison with the strongest published results. The robustness study also covers only independently injected noise. Because the core method is sound and the issues are correctable by re-running comparisons and qualifying claims, the contribution could be publishable after substantial revision.","major_comments":[{"comment":"The claim that S2DN 'surpasses the performance of state-of-the-art models' is contradicted by the paper's own reported numbers. In Table 9 (NELL-995 V4), RMPI achieves Hits@10 84.20 and MRR 72.63 versus S2DN's 74.89 and 64.12; in Table 1 (WN18RR V4), RMPI's Hits@10 82.33 exceeds S2DN's 82.15, and in Table 9 (NELL-995 V3) RMPI's MRR 80.07 exceeds S2DN's 76.22. The abstract and conclusion should either restrict the claim to 'most settings' or support an aggregated claim with a significance test over all twelve splits.","section":"Abstract; Conclusion; Tables 1 and 9"},{"comment":"The comparison does not establish superiority relative to the strongest published inductive KGC models because of a protocol mismatch. Appendix B.1.1 states that ranking is done against 50 randomly sampled negative entities, and Appendix B.2.3 excludes RED-GNN and AdaProp solely because they rank against all entities; GraIL is discussed and used in the visualization but does not appear in the main result tables. Under the chosen protocol the numbers may be internally consistent, but they are not directly comparable to published SOTA numbers from these baselines. The authors should either rerun the omitted baselines under a common protocol or explicitly state in the abstract and conclusion that the comparison excludes these methods.","section":"Appendix B.1.1; Appendix B.2.3; Tables 1, 2, and 9"},{"comment":"The manuscript states in Appendix B that each experiment was repeated five times, but all reported metrics are single numbers without standard deviations or significance tests. Several decisive comparisons are close (for example, WN18RR V4 Hits@10 82.15 versus 82.33, and NELL-995 V1 Hits@1 45.00 versus RMPI's 47.50), so without variance information the reader cannot judge whether the gains are meaningful. The authors should report mean and standard deviation over the five runs and, where appropriate, paired significance tests.","section":"Appendix B; Tables 1, 2, and 9"},{"comment":"The robustness claim is demonstrated only for independently injected noise. The structure refining module treats edges as mutually independent Bernoulli variables and scores reliability by the inner product of MLP-transformed node features, so correlated noise (for example, clusters of false facts among similar entities) is not covered by the experiments, which inject random relation replacements and random negative triples. If the robustness claim is meant to cover realistic contamination, the authors should either add experiments with correlated noise or restrict the claim to the tested noise model.","section":"Eqs. (5)-(6); 'Robustness of S2DN (RQ3)'; Tables 3, 4, and 8"},{"comment":"The information-bottleneck discussion is not tightly connected to the actual loss. Lemma 1 bounds I(\\tilde{E}; E_n), but the objective in Eq. (12) replaces mutual-information terms with a KL divergence between smoothed and original relation embeddings and a cross-entropy term; the paper does not show that minimizing Eq. (12) is equivalent to minimizing the bound in Eq. (10). This section should be labeled as intuition or made rigorous before being presented as a theoretical justification.","section":"Theoretical Discussion of Smoothing; Eq. (12)"}],"minor_comments":[{"comment":"The hyperparameter analysis says that S2DN achieves optimal performance with k = 4 for NELL-995 V1 and later designates NELL-995 as 4-hop, whereas Table 5/6 and Section B.1.2 set k = 2 for NELL-995; this inconsistency should be resolved.","section":"Appendix C.4.3; Table 5/6; Section B.1.2"},{"comment":"Tables 5 and 6 are duplicate 'hyperparameter details of S2DN' tables; the duplicate should be removed or renumbered.","section":"Appendix A/B; Tables 5 and 6"},{"comment":"The baseline TAGT is written as TACT in Appendix B.2.1 and B.2.2; please standardize the name throughout.","section":"Appendix B.2.1; Appendix B.2.2"},{"comment":"The robustness section contains dangling references such as 'as shown in Section ,' and the semantic-consistency paragraph uses '%' placeholders without a target; the section numbers need to be filled in.","section":"Section 'Robustness of S2DN (RQ3)'"},{"comment":"The phrase 'surpasses the performance of state-of-the-art models' should be qualified even after additional baselines are added, because the paper's own tables show that RMPI wins on several individual splits.","section":"Abstract; Conclusion"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the central method appears sound and the reported design choices are plausible, but the paper overclaims its empirical status. The decision should be contingent on the authors fixing the baseline comparison (including protocol alignment or explicit qualification) and reporting variance from the five runs. The appendix also needs cleanup of duplicated tables and contradictory hyperparameter statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper is worth a look for the method, not for the SOTA claim. S2DN combines Gumbel-softmax relation smoothing with learned Bernoulli edge pruning inside GraIL-style enclosing subgraphs, and the ablations show both pieces pull their weight. The robustness study, where the authors inject semantic and structural noise into the training KG, is the most useful part—it shows the denoising modules degrade more gracefully than RMPI across several noise levels. That is a genuine empirical contribution.\n\nWhere the paper hurts itself is in the headline comparison. AdaProp, RED-GNN, and GraIL are missing from the main tables, and the justification in Appendix B.2.3—that RED-GNN and AdaProp rank over all entities while GraIL-style methods use 50 sampled negatives—is exactly why the reported numbers can't be called SOTA. The paper's own Table 9 shows RMPI beating S2DN on NELL-995 V4 by a wide margin, and WN18RR V4 Hits@10 also favors RMPI. So \"surpasses state-of-the-art models\" is simply not established. The evaluation protocol is legitimate, but then the claim must be compared against baselines run under the same protocol.\n\nOther soft spots are minor: no error bars despite five runs; duplicate hyperparameter tables (Tables 5 and 6); a placeholder \"Section\" reference; and the Bernoulli edge-independence assumption could be problematic if noise is correlated, but that's a modeling choice, not a fatal flaw. The information-bottleneck lemma is a standard inequality dressed up; it doesn't add much.\n\nThe work is honest in its own terms: the modules are clearly described, code is available, and the authors openly state why some baselines were excluded. I don't see circularity in the reported numbers.\n\nWho gets value from this: researchers working on inductive KGC or graph structure learning, particularly on denoising subgraph reasoning. It deserves a serious referee, but only after the authors fix the comparison and soften the SOTA wording. If it were a submission in front of me, I'd recommend major revision, not rejection. I would not cite it as evidence of state-of-the-art performance as it stands.","headline":"A sensible denoising method for inductive KGC, but the SOTA claim is not supported by the reported experiments; both modules ablate well, and the robustness study is useful.","tokens_in":23232,"tokens_out":2595,"would_cite":false,"duration_ms":23187,"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":"This paper claims that inductive knowledge graph completion can be made more accurate by smoothing semantically similar relations into shared representations and by pruning unreliable edges from the local subgraph, and reports that S2DN…","keywords":["inductive knowledge graph completion","semantic smoothing","structure refining","knowledge graph denoising","subgraph reasoning","Gumbel-softmax","edge reliability","noisy knowledge graphs"],"falsifier":"Build a poisoned version of WN18RR V1 where the injected false edges deliberately connect nodes whose MLP-transformed features are similar (for example, by adding noise that targets nodes on the same shortest-path distance labels), and compare S2DN to its own no-structure-refining variant: if the full model fails to beat the variant, the feature-similarity reliability assumption is refuted.","tokens_in":22262,"feed_emoji":"🔗","tokens_out":7946,"duration_ms":63191,"temperature":0.7,"pith_summary":"Inductive knowledge graph completion asks a model to predict missing facts about entities it has never seen during training. The paper claims that the two biggest obstacles in this setting are inconsistent representations of semantically similar relations and noisy edges that surround newly emerged entities, and that both can be fixed by a single network that denoises meaning and structure together. The proposed S2DN smooths relation embeddings so similar relations share one semantic representation, while a structure-refining module scores every edge in the local subgraph and drops the ones that look unreliable. On WN18RR, FB15k-237, and NELL-995, the paper reports that S2DN outperforms existing rule-based and GNN-based inductive KGC methods, and that it degrades less than baselines when false facts are deliberately injected. If true, this would mean that explicit denoising of relations and edges is a practical route to robust reasoning about unseen entities, not just a stronger encoder.","feed_headline":"Smoothing relations and pruning noisy edges improves KG completion","feed_subtitle":"S2DN keeps similar relations consistent and drops unreliable edges, reporting higher Hits@10 on three benchmarks.","key_machinery":"The load-bearing mechanism is a pair of denoising modules applied to the enclosing subgraph of a candidate triple. The semantic smoothing module learns a categorical assignment of each observed relation to a smaller set of semantic clusters via Gumbel-softmax reparameterization, blurring relations like 'located in' and 'lie in' into one representation; an information-bottleneck style argument and a KL penalty justify keeping the smoothed embedding close to the original. The structure refining module computes an edge-reliability score as the sigmoid of the inner product of MLP-transformed node-feature vectors, models all edges as independent Bernoulli variables, relaxes them with the concrete distribution for differentiability, and removes edges whose probability falls below 0.5. The smoothed graph is encoded with a relational GNN and the refined graph with a GCN, and their pooled representations are concatenated and fed to an MLP classifier. This machinery is what lets the model discard unreliable interactions while preserving relation semantics for unseen entities.","core_discovery":"The paper's central claim is that 'unconvincing knowledge' in inductive knowledge graph completion has two independent sources, semantic inconsistency and structural noise, and that a network addressing both can beat the state of the art. Semantic smoothing maps relations with similar meanings into a shared embedding using a Gumbel-softmax over the relation set, with a KL term keeping the smoothed representation from drifting too far from the original. Structure refining estimates a reliability probability for each edge in the enclosing subgraph from the inner product of MLP-transformed node features, treats edges as independent Bernoulli variables, and prunes edges with reliability below 0.5 before applying a GCN. The smoothed subgraph is encoded by a relational GNN and the refined subgraph by a GCN, and the two global readouts are concatenated and classified. The paper reports this design outperforms the compared baselines on the three benchmark datasets and that ablations removing either module cause clear drops in Hits@1, Hits@10, and MRR.","pith_inferences":["A testable extension the authors do not run: apply the same two modules to transductive KGC benchmarks, where entities are seen during training; if smoothing relations still helps there, the benefit is not specific to inductive generalization but to noise generally.","An adversarial test the paper does not run: insert false edges specifically between nodes with similar learned features, since the structure refiner assumes feature-similar connected nodes are trustworthy; correlated noise of this kind may defeat the pruning.","The smoothed relation assignment learned by the network could be read as an emergent relation ontology; comparing it against human-curated relation clusters on FB15k-237 would be a direct external check of the semantic-consistency claim.","For the biomedical application the paper names as future work, the natural next experiment is to poison a drug-interaction or protein-function KG with literature-derived false associations and check whether the same robustness pattern holds when relations are more numerous and heterogeneous."],"forward_implications":["The performance gains reported over rule-based and GNN baselines imply that inductive KGC systems do not need cleaner graphs; a network that smooths relations and prunes edges before reasoning can recover much of the lost accuracy.","The contamination experiments imply that filtering edges by learned node-feature similarity is an effective defense against randomly injected false facts, both semantic replacements and randomly sampled triples.","The ablation results imply that neither module alone is sufficient: removing semantic smoothing hurts most on semantically noisy graphs, while removing structure refining hurts most on structurally noisy graphs.","Because relation smoothing is learned from context rather than fixed vocabularies, the method is designed to transfer to new relation sets, which is exactly what inductive completion requires."],"supporting_citations":[{"why":"Defines enclosing-subgraph reasoning and the inductive splits and node features that S2DN inherits.","marker":"Teru, Denis, and Hamilton 2020"},{"why":"Shows relational path information helps inductive KGC and provides one of the GNN baselines S2DN compares against.","marker":"Xu et al. 2022"},{"why":"Provides the relational message-passing baseline that S2DN must beat in both clean and noisy settings.","marker":"Geng et al. 2023"},{"why":"Supplies the Gumbel-softmax reparameterization that makes semantic smoothing differentiable.","marker":"Jang, Gu, and Poole 2017"},{"why":"Supplies the GCN used to encode the structure-refined subgraph.","marker":"Kipf and Welling 2017"},{"why":"Supplies the relational graph convolutional layer used on the semantically smoothed subgraph.","marker":"Schlichtkrull et al. 2018"}],"fun_headline_variants":["S2DN: Smooth semantics, prune noisy edges for inductive KGC","S2DN: Clean semantics and edges for better KG completion","S2DN: Fixing both semantic and structural noise in KGs","S2DN: Semantic smoothing and edge pruning for KGC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that how reliable an edge is can be judged from the similarity of learned node-feature vectors, and that edges are independent, so a false fact connecting two feature-similar nodes will survive pruning and can mislead the prediction.","fun_headline_variants_meta":{"raw":{"variants":["S2DN: Smooth semantics, prune noisy edges for inductive KGC","S2DN: Clean semantics and edges for better KG completion","S2DN: Fixing both semantic and structural noise in KGs","S2DN: Semantic smoothing and edge pruning for KGC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001218,"raw_usage":{"total_tokens":5018,"prompt_tokens":963,"completion_tokens":4055,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":3980}},"tokens_in":579,"tokens_out":4055,"duration_ms":23032,"temperature":1.0,"reasoning_tokens":3980,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:03:22.523730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a poisoned version of WN18RR V1 where the injected false edges deliberately connect nodes whose MLP-transformed features are similar (for example, by adding noise that targets nodes on the same shortest-path distance labels), and compare S2DN to its own no-structure-refining variant: if the full model fails to beat the variant, the feature-similarity reliability assumption is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines enclosing-subgraph reasoning and the inductive splits and node features that S2DN inherits."},{"cited_title":"Z.; Chen, M.; Jiang, S.; Zhang, W.; and Chen, H","cited_arxiv_id":null,"evidence_quote":"Provides the relational message-passing baseline that S2DN must beat in both clean and noisy settings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gumbel-softmax reparameterization that makes semantic smoothing differentiable."},{"cited_title":"N.; and Welling, M","cited_arxiv_id":null,"evidence_quote":"Supplies the GCN used to encode the structure-refined subgraph."},{"cited_title":"N.; Bloem, P.; Van Den Berg, R.; Titov, I.; and Welling, M","cited_arxiv_id":null,"evidence_quote":"Supplies the relational graph convolutional layer used on the semantically smoothed subgraph."}],"review_version":1}