{"id":"31b96630-b14c-43af-b1f3-fe5c13acc07f","arxiv_id":"2505.12894","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HyperDet detects rumor sources in hypergraphs by combining static topology, dynamic interaction hyperedges, autoencoder feature enrichment, and multi-head attention, outperforming pairwise baselines on eight simulated datasets.","lead":"This paper introduces HyperDet, a neural network method that detects the original spreaders of a rumor in social networks modeled as hypergraphs, where groups interact at once rather than just in pairs. It builds richer node features and attention weights, and reports higher accuracy than existing pairwise methods on simulated benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact infection timestamps in the input features may trivialize source detection; the reported SOTA gains could reflect feature leakage rather than higher-order modeling, and no timestamp-only baseline is reported.","rationale":"The empirical SOTA claim is the central contribution. My concern is that the evaluation may be internally confounded: the model is given exact infection timestamps (Eq. 5), which in the synthetic IC model trivially identify sources. The paper never reports a timestamp-only baseline or states whether baselines received the same feature. This does not necessarily mean the method is wrong, but it means the current evidence does not establish that the gains come from hypergraph-aware modeling. The reader focused on external validity of the synthetic model; I focus on internal validity. A simple ranking experiment would settle the issue. Until then, the correct disposition is conditional: require the authors to add the timestamp baseline, remove or noise the timestamp feature in an ablation, and release code/data for reproducibility.","tokens_in":13875,"tokens_out":10253,"duration_ms":111753,"concrete_test":"Run the published experimental pipeline (or request code) and add a baseline that ranks spreader nodes by X_i^2 (smallest first) and predicts the top |s| nodes; compute F1/AUC on the same 8:2 splits. In addition, retrain HyperDet with the timestamp component of X_i removed, and give baselines the same timestamp feature. If the timestamp-only ranker matches or exceeds HyperDet, or if removing X_i^2 collapses HyperDet's advantage, the central SOTA claim is confounded by feature leakage rather than higher-order modeling.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is internal to the evaluation. In the synthetic IC model (Section 5.1), 5% of nodes are chosen as sources and become spreaders at t=0; all other infections occur at later steps. The raw feature X_i^2 defined in Eq. (5) is exactly the infection timestamp t_i for every spreader node and -1 for ignorant nodes. Therefore the source set is precisely the set of spreader nodes with the minimal X_i^2 value. A trivial ranker over this single feature recovers the ground truth (up to ties among sources) without any hypergraph modeling. The paper does not state whether baselines were given this timestamp feature, does not report a timestamp-only baseline, and does not explain why HyperDet's F1 is 0.63-0.80 and AUC 0.86-0.92 instead of near 1.0 if the timestamp is exact. If timestamps were hidden, noised, or unavailable, that must be stated; if they were available to HyperDet but not to baselines, the comparison is unfair. Either way, the claimed 8-15% F1 improvement over propagation-based methods would not demonstrate the value of the hypergraph relationship construction or attention fusion.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HyperDet, a supervised method for source detection in hypergraphs. It builds an Interactive Relationship Construction (IRC) module that combines the static hypergraph topology with dynamic state-based hyperedges, and a Feature-rich Attention Fusion (FAF) module that applies a hypergraph autoencoder and multi-head attention to learned node features, together with a class-balanced loss. The method is evaluated on eight datasets under a synthetic hypergraph independent cascade model, reporting consistent gains over centrality-, user-state-, and propagation-based baselines, along with ablations, early-detection analyses, and robustness checks under alternative diffusion models.","tokens_in":14128,"tokens_out":7898,"duration_ms":84228,"significance":"If the evaluation is valid, HyperDet would be a meaningful contribution: it defines a hypergraph-specific source-detection setting, and the reported F1 improvements of roughly 8-15% over propagation-based baselines are substantial. The paper also includes a broad set of datasets, module ablations, and sensitivity-style experiments. However, the central claim is not currently established because the raw input features appear to contain exact infection timestamps that make the synthetic source-detection task nearly trivial, and because the evaluation lacks repeated trials, error bars, and statistical testing. The paper does not release code, which further limits reproducibility. The strength of the contribution therefore depends on resolving the timestamp-leakage concern.","major_comments":[{"comment":"The raw feature X_i^2 defined in Eq. (5) is the exact infection timestamp t_i for every spreader node and -1 for ignorant nodes. In the synthetic IC model of Section 5.1, 5% of nodes are selected as sources and become spreaders at t=0, while all other infections occur at later steps. Consequently, the set {v_i : X_i^2 = 0} is exactly the ground-truth source set, and a trivial threshold or ranker on this single feature solves the task perfectly without any hypergraph modeling, giving F1 and AUC near 1.0. The paper does not report a timestamp-only baseline, does not state whether the baselines were given the same timestamp feature, and does not explain why HyperDet's reported F1 is only 0.63-0.80 and AUC 0.86-0.92 under these conditions. If timestamps were available to HyperDet but not to baselines, the comparison is unfair; if they were hidden, noised, or unavailable, that must be stated explicitly. Please remove the exact timestamp from the input, use noised or censored timestamps, add a timestamp-only or minimum-timestamp baseline, and report all methods under identical feature availability. This is necessary before the claimed 8-15% improvement can be attributed to higher-order relationship modeling rather than to feature leakage.","section":"Section 4.1, Eq. (5); Section 5.1"},{"comment":"All results in Table 1 appear to be single runs with no error bars, no standard deviations, and no significance tests. The evaluation also uses one fixed synthetic propagation model (p ~ U(0,0.5), p_delta = 0.3(|e ∩ G+|/|e|), 5% sources, 30% snapshot threshold), and the datasets are not actual social networks but standard UCI/benchmark datasets with the synthetic process superimposed. Under this protocol, the reported performance gaps between methods may reflect run-to-run variability rather than model quality, and the conclusions cannot be extrapolated to other propagation dynamics. Please report means and standard deviations over multiple random seeds, include a statistical significance test (e.g., paired t-test or Wilcoxon signed-rank test), and provide sensitivity analyses over the propagation model parameters. The alternative-model results in Table 4 should also be compared with baselines, not only with HyperDet.","section":"Section 5.1, Table 1"}],"minor_comments":[{"comment":"The positional feature X_i^3 is a k-dimensional vector for spreader nodes but is written as the scalar -1 for ignorant nodes; if the concatenated raw feature is to have a fixed dimension, the -1 case needs to be defined as a vector of -1s of the same length, or otherwise standardized.","section":"Section 4.1, Eq. (8)"},{"comment":"There is a dimension mismatch in the attention coefficient formula: W is defined in R^{l'_w x l_w}, but it is applied to the concatenated vector [X_i || theta_j], which has length 2l_w; the dimensions of W or of the concatenation need to be corrected for the model to be implementable as written.","section":"Section 4.2, Eq. (12)"},{"comment":"The sentence 'w/o A achieves the best results' is ambiguous because the table shows that the full HyperDet model outperforms w/o A; please clarify that w/o A is the best among the three attention variants (w/o A, w/ AL, w/ AS) and that the full attention fusion still contributes positively.","section":"Section 5.3, Table 3"},{"comment":"The experimental settings refer to 'small networks (G1-G4)', but the datasets are named Zoo, House, NTU2012, and Mushroom; please align the notation with the dataset names used in Table 1.","section":"Section 5.1"},{"comment":"The claim 'We are the first to formalize the problem of source detection in hypergraphs' should be softened or supported by a more thorough literature search, since the paper does not discuss any prior hypergraph-based source localization work and the current survey in Section 2 stops at hypergraph representation learning.","section":"Section 1, Contributions"},{"comment":"Table 4 reports HyperDet under SI, SIS, SIR, and IC models, but it does not include any baseline method under the same alternative models; adding at least the strongest baseline (e.g., GIN-SD) would make the robustness claim more convincing.","section":"Section 5.2, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The timestamp-leakage issue in Eq. (5) and the synthetic evaluation protocol are serious enough that the current version should not be accepted without a re-evaluation. The authors should verify the 'first' novelty claim against the broader hypergraph source-localization literature, and I would encourage them to release code and data splits for reproducibility. The self-citation pattern is noticeable but not improper; comparing against their own GIN-SD is a legitimate benchmark choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paper to pose source detection on hypergraphs, and I believe that claim. The IRC module does something real: it adds dynamic interaction hyperedges for spreader and ignorant sets to the static incidence matrix, and the ablation shows both contribute. The FAF module is standard hypergraph attention plus an autoencoder, but it's put together competently. The gains over GIN-SD and IVGD are large and consistent across eight datasets. That part is worth a look.\n\nThe soft spots are not minor. The biggest one, which I'd want a referee to chase hard, is the timestamp feature. Eq. (5) defines X_i^2 = t_i for every spreader and -1 otherwise, and in the synthetic IC setting the sources are exactly the spreaders with t=0. That means a one-line baseline that picks the nodes with the smallest timestamp recovers the source set, up to ties. The paper does not report such a baseline, and does not say whether the timestamps were hidden, noised, or made unavailable to the baselines. HyperDet's own F1 scores are 0.63-0.80, not near 1.0, which makes me think either the implementation does not actually use exact t_i or the model is failing to exploit a feature that should make the task trivial. Either way, the main empirical claim does not currently demonstrate the value of higher-order modeling.\n\nThere are secondary issues. No code or data released, no error bars or significance tests, and the whole evaluation is on synthetic cascades with one noise model; Table 4 shows robustness across SI/SIS/SIR but only for HyperDet, not against baselines. The class balancing, the early-detection and incompleteness plots are fine, but they don't fix the leakage question. The citation pattern is okay; GIN-SD is one of the authors' own, but it is a sensible baseline, not a circular one.\n\nBottom line: the paper deserves a serious referee, not a desk reject, because the problem is new and the architecture is reasonable. But the reviewer must demand (a) a statement of what exactly the timestamp feature contains in the experiments, (b) a timestamp-only baseline, and (c) either code or a much more detailed protocol. If the leakage is confirmed, the SOTA claims in Table 1 don't stand. If the timestamps are noised or hidden and the authors show it, this could be a decent contribution.","headline":"The first hypergraph source-detection paper with a usable architecture, but the evaluation may leak the answer through exact infection timestamps in the input features.","tokens_in":14644,"tokens_out":3308,"would_cite":false,"duration_ms":35821,"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 reports the first source-detection method designed for hypergraphs, with F1 gains of 8–15 percent over propagation-aware baselines and 23–30 percent over centrality/user-state methods across eight datasets.","keywords":["hypergraph","rumor source detection","source localization","higher-order interactions","graph neural networks","attention mechanism","information propagation","independent cascade model"],"falsifier":"Simulate hypergraph cascades with group-propagation probability set to zero ($p_\\Delta = 0$) while keeping every other setting identical, and compare HyperDet against GIN-SD; if the reported 8–15 percent F1 advantage disappears or reverses, the paper's attribution of the gains to higher-order interaction modeling is refuted.","tokens_in":13697,"feed_emoji":"📢","tokens_out":12633,"duration_ms":119808,"temperature":0.7,"pith_summary":"Rumor source detection has mostly assumed information spreads through pairwise contacts, but in classrooms, households, and online groups people influence one another in larger clusters. This paper claims to be the first to formalize source detection on hypergraphs—networks whose edges can contain many users at once—and proposes HyperDet to solve it. HyperDet builds a node-relationship representation from both the static hypergraph structure and the live split between spreaders and ignorants, then learns node representations through an autoencoder and a multi-head attention mechanism that emphasizes nodes with strong transmission roles. On eight datasets with simulated hypergraph cascades, the paper reports F1-score gains of 8–15 percent over propagation-aware baselines and 23–30 percent over centrality- and user-state-based baselines. If those numbers hold, group structure is not something to flatten into pairwise edges but a resource for pinpointing where a rumor began.","feed_headline":"HyperDet finds rumor sources in hypergraphs, with F1 up 8–15%","feed_subtitle":"Group influence, not just pairwise contacts, carries the signal; the model wins across eight datasets.","key_machinery":"The load-bearing construction is the augmented incidence matrix $$H' = H \\parallel H_{\\mathrm{ig}} \\parallel H_{\\mathrm{sp}},$$ which folds three kinds of relationship into one hypergraph: the original static hyperedges, one hyperedge containing every current spreader, and one containing every ignorant node. The raw node feature $X_i$ concatenates a state label (+1 for spreader, −1 otherwise), the propagation timestamp (or −1 for ignorants), and the $k$ smallest nontrivial Laplacian eigenvectors of the infected subgraph as positional encoding. These inputs feed the Feature-rich Attention Fusion module: a hypergraph autoencoder whose layers alternate node-to-hyperedge and hyperedge-to-node attention convolution, followed by multi-head self-attention that gives different nodes different weight, with the loss balanced by $\\rho = |s|/(n - |s|)$. The whole pipeline is what lets the model learn node representations under higher-order relations instead of pairwise flattening.","core_discovery":"The central claim, stated the way the authors would state it, is that a source detector which models higher-order interactions directly, rather than flattening them into pairwise edges, can locate rumor origins in hypergraphs more accurately than pairwise methods. HyperDet takes a single snapshot of an evolving hypergraph: its topology, each node's spreader/ignorant state, and propagation timestamps, and outputs a predicted source set. The Interactive Relationship Construction module concatenates the static incidence matrix $H$ with two dynamic hyperedges $H_{\\mathrm{ig}}$ (all ignorants) and $H_{\\mathrm{sp}}$ (all spreaders), and composes raw features from state, timestamp, and Laplacian positional encoding of the infected subgraph. The Feature-rich Attention Fusion module autoencodes these features through hypergraph attention convolution, reweights nodes by multi-head self-attention, and uses a class-balancing coefficient to counter the rarity of sources. Across the authors' experiments on eight datasets, HyperDet achieves the best accuracy, F1-score, and AUC on every dataset.","pith_inferences":["Beyond the paper: the decisive test is real-world deployment. Because all reported cascades are simulated with one specific group-pressure formula, we should expect the 8–15 percent edge to shrink or shift when applied to real message-forwarding data with ground-truth origins.","If higher-order interaction modeling is truly the cause of the gains, then sweeping the group-propagation probability $p_\\Delta$ downward from 0.3 toward 0 should make HyperDet's advantage over pairwise baselines shrink monotonically; this is a cheap, direct check of the paper's main explanatory story.","The idea of encoding the current spreader/ignorant partition as two extra hyperedges is portable. It could be applied to other dynamic hypergraph tasks—for instance, predicting which node will be affected next, or detecting coordinated groups—where the current infection state is a conditioning variable rather than the target."],"forward_implications":["Pairwise source detectors cannot simply be copied onto hypergraphs: the paper shows their performance degrades when group interactions are present, whereas a detector that models the group structure directly keeps high F1.","Early detection works: as the rumor scale grows from 10% to 30%, HyperDet remains ahead of all baselines on every dataset, so the model does not need a saturated network to identify the source.","It tolerates missing data: when up to 25% of nodes lack features, HyperDet's F1 declines less than the baselines, because structural positional encoding still carries information.","The architecture is not tied to one diffusion model: stable results are reported under SI, SIS, SIR, and independent-cascade dynamics, so the two-module design generalizes across propagation assumptions."],"supporting_citations":[{"why":"Introduces source centrality for rumor-source identification; the classic formulation that HyperDet extends and the basis of centrality baselines.","marker":"[Shah and Zaman, 2011]"},{"why":"Supplies the higher-order interactions perspective and hypergraph contagion models that motivate the problem.","marker":"[Battiston et al., 2020]"},{"why":"Defines hypergraph neural network message passing (node-to-hyperedge, hyperedge-to-node) that the FAF autoencoder builds on.","marker":"[Feng et al., 2019]"},{"why":"Provides the graph attention mechanism that HyperDet adapts into multi-head attention over hyperedges.","marker":"[Veličković et al., 2017]"},{"why":"IVGD is a propagation-information baseline against which HyperDet claims 8–15 percent F1 gains.","marker":"[Wang et al., 2022]"},{"why":"SL-VAE is a propagation-aware source-localization baseline that infers graph diffusion patterns.","marker":"[Ling et al., 2022]"},{"why":"GIN-SD, the strongest pairwise baseline, uses positional encoding and attentive fusion on graphs with incomplete nodes; HyperDet extends this idea to hypergraphs.","marker":"[Cheng et al., 2024b]"},{"why":"Supplies the House hypergraph dataset and generative hypergraph clustering methodology used in evaluation.","marker":"[Chodrow et al., 2021]"}],"fun_headline_variants":["HyperDet spots rumor sources via hypergraph group dynamics","HyperDet outperforms pairwise models by modeling higher-order interactions","Higher-order hypergraph model finds rumor sources with best F1","HyperDet models group interactions, beating pairwise on 8 datasets","HyperDet tops F1 on 8 datasets via hypergraph group modeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic hypergraph cascade used for evaluation—with individual spread probability drawn uniformly from $[0, 0.5]$ and group pressure set to $p_\\Delta = 0.3\\,(|e \\cap G^+|/|e|)$—captures the way rumors actually move through real group interactions.","fun_headline_variants_meta":{"raw":{"variants":["HyperDet spots rumor sources via hypergraph group dynamics","HyperDet outperforms pairwise models by modeling higher-order interactions","Higher-order hypergraph model finds rumor sources with best F1","HyperDet models group interactions, beating pairwise on 8 datasets","HyperDet tops F1 on 8 datasets via hypergraph group modeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001332,"raw_usage":{"total_tokens":5397,"prompt_tokens":902,"completion_tokens":4495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":4410}},"tokens_in":518,"tokens_out":4495,"duration_ms":31258,"temperature":1.0,"reasoning_tokens":4410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:00.803355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate hypergraph cascades with group-propagation probability set to zero ($p_\\Delta = 0$) while keeping every other setting identical, and compare HyperDet against GIN-SD; if the reported 8–15 percent F1 advantage disappears or reverses, the paper's attribution of the gains to higher-order interaction modeling is refuted.","supporting_citations":[{"cited_title":"Rumors in a network: Who’s the culprit?IEEE Transac- tions on Information Theory, 57(8):5163–5181,","cited_arxiv_id":null,"evidence_quote":"Introduces source centrality for rumor-source identification; the classic formulation that HyperDet extends and the basis of centrality baselines."},{"cited_title":"Networks beyond pairwise interactions: Structure and dynamics","cited_arxiv_id":null,"evidence_quote":"Supplies the higher-order interactions perspective and hypergraph contagion models that motivate the problem."},{"cited_title":"Hypergraph neural networks","cited_arxiv_id":null,"evidence_quote":"Defines hypergraph neural network message passing (node-to-hyperedge, hyperedge-to-node) that the FAF autoencoder builds on."},{"cited_title":"An invertible graph diffusion neural network for source localization","cited_arxiv_id":null,"evidence_quote":"IVGD is a propagation-information baseline against which HyperDet claims 8–15 percent F1 gains."},{"cited_title":"Source localization of graph diffusion via variational autoencoders for graph inverse problems","cited_arxiv_id":null,"evidence_quote":"SL-VAE is a propagation-aware source-localization baseline that infers graph diffusion patterns."},{"cited_title":"Generative hypergraph clustering: From blockmodels to modularity.Science Advances, 7(28):eabh1303,","cited_arxiv_id":null,"evidence_quote":"Supplies the House hypergraph dataset and generative hypergraph clustering methodology used in evaluation."}],"review_version":1}