{"id":"04066e29-cc3a-4888-a035-a65ca507f0fa","arxiv_id":"2412.19227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DHy-MFND learns news embeddings from text, propagation trees, and a dynamically refined hypergraph, and reports state-of-the-art accuracy on PolitiFact and Gossipcop.","lead":"This paper introduces DHy-MFND, a fake news detection model that combines text, propagation tree, and hypergraph views, and dynamically refines the hypergraph structure during training. On the PolitiFact and Gossipcop datasets, it reports the highest accuracy and F1 among nine baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contrastive loss in Eq. (16) is written over all N news pieces using true labels; with a 6:2:2 split this can leak validation/test labels into training unless an unstated train mask is applied. This must be resolved before the Table 3 superiority claim is credible.","rationale":"The reader's weakest_assumption is exactly the concern I would stress-test: Eq. (16) is written over all N news pieces and uses ground-truth labels to define positive and negative pairs, while the paper never states that only training labels are used. If the loss sees validation or test labels, the contrastive objective encodes test class membership into the learned embeddings, which would inflate the reported accuracy and F1. This is load-bearing because the central claim is empirical superiority over baselines in Table 3. The test I propose is concrete: inspect the code or rerun with a training mask. If the mask exists, the concern is resolved; if not, the reported results are invalid. I agree with the reader's conditional verdict because the issue is serious but addressable, and I do not see a stronger independent objection. The paper does have independent support in its careful ablation structure and the use of standard benchmark datasets, but those do not mitigate the label-leakage ambiguity. No ad hominem is intended; the issue is a notation/implementation gap that must be closed before the claim is accepted.","tokens_in":12146,"tokens_out":3136,"duration_ms":35212,"concrete_test":"Obtain the released code, or ask the authors for it, and check whether the computation of Eq. (16) in Algorithm 1 applies a training-index mask to the label sets K(i) and T(i). Then rerun the full experiments on PolitiFact and Gossipcop with the contrastive loss explicitly restricted to training-set nodes (and with cross-entropy masked to training labels), using the same hyperparameters and 6:2:2 split. If the resulting Acc/F1 values drop by more than one standard deviation relative to Table 3, or if the code shows that all labels were used in Eq. (16), the label-leakage concern lands and the empirical claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on the numbers in Table 3, so any mechanism that lets the model see test labels during training would invalidate the comparison. The paper defines the contrastive loss in Eq. (16) as a sum over all N news pieces, with positive pairs K(i) = {k : y_k = y_i} and negative pairs T(i) = {t : y_t ≠ y_i}, where y are ground-truth labels. Algorithm 1 takes all labels Y as input and never restricts the contrastive loss to the training split. The dataset is split 6:2:2, but the hypergraph and propagation trees are constructed for all news pieces, so if Eq. (16) is evaluated over all N, the model's training objective directly uses validation/test labels to pull same-class embeddings together and push different-class embeddings apart. That is label leakage, not mere transductive use of unlabeled test features. The same all-N notation appears in the classification loss in Eq. (15), though that could be a batch-size convention; however, for Eq. (16) the paper never states that K(i) and T(i) are restricted to the training set. The reader's concern is therefore well-founded: without a stated or implemented mask, the reported 92.81% Acc on PolitiFact and 98.84% Acc on Gossipcop could be inflated by the contrastive objective seeing test labels. This is an addressable ambiguity rather than a confirmed flaw, but it is the most load-bearing uncertainty because it directly determines whether the headline results are valid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DHy-MFND, a multi-view fake news detection model that combines BERT-based text embeddings, a GraphSAGE propagation-tree encoder, and an attention-based hypergraph neural network, together with a dynamic hypergraph structure learning (DHSL) module that refines the incidence matrix during training. A supervised InfoNCE contrastive loss aligns propagation-tree and hypergraph embeddings for news with the same authenticity label. Experiments on PolitiFact and Gossipcop report accuracy and F1 scores over nine baselines, with the proposed model achieving the highest mean scores in Table 3. The paper also includes an ablation study, a runtime table, and a sensitivity analysis of the threshold ratio p_thd.","tokens_in":12508,"tokens_out":7040,"duration_ms":61344,"significance":"If the reported results are valid, the paper makes a modest but useful contribution by showing that combining text, propagation-tree, and hypergraph views with dynamic hypergraph structure learning and supervised contrastive learning can improve fake news detection on two standard benchmarks. The paper is generally reproducible in structure: it provides algorithm pseudocode, reports means and standard deviations for the main comparisons, and includes a hyperparameter sensitivity analysis. The central empirical claim, however, hinges on the correct isolation of training labels in the contrastive loss and on the statistical robustness of the reported gains, both of which need clarification.","major_comments":[{"comment":"The concern raised in the stress-test about label leakage is well-founded. The contrastive loss in Eq. (16) is written as a sum over all N news pieces, with K(i) = {k : y_k = y_i} and T(i) = {t : y_t ≠ y_i} using ground-truth labels. The dataset is split 6:2:2, but the hypergraph and all computations are described on the full set of N nodes. Algorithm 1 takes the full label vector Y as input and computes the contrastive loss without any visible restriction to the training split. If validation or test labels participate in the positive/negative pair construction during training, the reported Acc/F1 numbers in Table 3 are invalid. Please state explicitly whether a training mask is applied; if not, rerun the experiments with the contrastive loss computed only on training-set nodes and report whether the headline results still hold.","section":"Eq. (16), Algorithm 1"},{"comment":"The cross-entropy loss in Eq. (15) is given as L_ce = -1/N Σ Σ ŷ_{i,c} log(y_{i,c}), with y_{i,c} defined as the ground-truth label and ŷ_{i,c} as the predicted probability. As written, the ground-truth label appears inside the logarithm, which is not a valid classification loss and would involve log(0) for the zero class. Please correct the equation to the standard form (ground truth outside the log, predicted probability inside) and confirm that the implementation matches the corrected form, since Eq. (17) combines this loss with the contrastive loss.","section":"Eq. (15)"},{"comment":"The claim that DHy-MFND 'consistently outperforms' the baselines is not fully supported by the reported variance. On PolitiFact, the mean accuracy gain over FinerFact is 92.81±2.50 vs 91.48±1.89, a difference within one standard deviation, and no significance tests are reported. On Gossipcop, the gain over HGFND is 98.84±0.30 vs 97.46±0.30, which is more substantial, but the overall claim needs statistical backing. Please report paired significance tests (e.g., t-test or Wilcoxon signed-rank over the 20 runs) and provide standard deviations for the ablation results in Table 5, which are currently reported without any variance.","section":"Table 3"}],"minor_comments":[{"comment":"In Eq. (14) the reconstructed incidence matrix is referred to as both H_re and H_rec, and Algorithm 1 calls 'Generate H_re using Eq. (13)' twice; please disambiguate the update of H^{(l+1)} in the DHSL recursion.","section":"Eq. (14), Algorithm 1"},{"comment":"The sensitivity plot in Figure 4 does not label the y-axis; please state which metric (Acc or F1) is shown and make the axis labels explicit.","section":"Figure 4"},{"comment":"The paper says 9 baselines, but Table 3 lists three UPFD variants (GCN, GAT, SAGE) in addition to the other methods; please clarify whether UPFD is counted as one or three separate baselines in the comparison.","section":"Baselines"},{"comment":"The dataset name is spelled 'Politifact' in Table 2 and 'PolitiFact' in Table 3 and elsewhere; please standardize the spelling.","section":"Table 2 and Table 3"},{"comment":"The paper does not describe how the validation split is used (e.g., early stopping or hyperparameter selection); please specify this to rule out any selection of the best epoch or test configuration on the test set.","section":"Experimental Setup"},{"comment":"In the Hyperparameter Analysis section, the sentence 'For 0 ≤ p_thd ≤ 1, we conduct repeated experiments...' appears twice; please remove the duplicate.","section":"Hyperparameter Analysis"}],"recommendation":"major_revision","confidential_remarks":"The main unresolved issue is the potential label leakage in the contrastive loss. If the authors confirm that a training mask is applied to Eq. (16), the paper could be publishable after the statistical and presentation revisions. I would encourage requesting code or a detailed data-preprocessing description to verify the mask, and to confirm that the corrected cross-entropy loss in Eq. (15) matches the implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean engineering combination: text (BERT), propagation tree (GraphSAGE), and hypergraph (attention-based HGNN) views, with dynamic hypergraph structure learning borrowed from HSL and HGFND's hypergraph construction. The idea of optimizing the hypergraph while learning embeddings, plus InfoNCE between propagation and hypergraph views, is coherent and worth testing. The paper is clearly written and the architecture is reproducible in principle. It does not break new conceptual ground, but it is a reasonable step forward for a crowded subfield.\n\nThe main problem is the contrastive loss in Eq. (16). It is written as a sum over all N news pieces, with positive and negative sets defined by ground-truth labels. Algorithm 1 takes all labels as input and never restricts the loss to the training split. With a 6:2:2 split and a hypergraph built over all news nodes, this can leak validation/test labels into training. The paper never states that K(i) and T(i) are masked to training indices. That is not a minor typo; it directly affects whether Table 3's numbers mean anything. The stress-test note is right to flag it as the load-bearing uncertainty.\n\nOther soft spots are smaller. The ablation study (Table 5) reports no variance, even though the experiment section says all models were run 20 times and means with standard deviations are reported. That makes it hard to judge whether differences like w/o HG vs w/o Pro are real. The PolitiFact dataset has only 314 news pieces, and the claimed gain over FinerFact (92.81±2.50 vs 91.48±1.89) is within one standard deviation; the Gossipcop gain over HGFND is more convincing. Hyperparameter analysis covers only p_thd, not lambda or tau, and the chosen values for those are not given.\n\nThe paper cites HSL and HGFND properly, and the combination is new enough for an incremental paper. The central idea is not flawed; the execution has a reproducibility gap that can be fixed by rewriting the loss with an explicit training mask and reporting ablations with variance.\n\nI would not cite the numbers as they stand, but I would bring this to a reading group as a case study in how transductive settings can blur label leakage in contrastive learning. A serious editor should send it to peer review, asking the authors to clarify the masking and provide code/data. If the leakage is resolved, this is a publishable incremental contribution.","headline":"A sensible three-view fake news detector whose headline numbers rest on an unstated train/test mask in the contrastive loss; fix that and the paper is a solid incremental contribution.","tokens_in":12982,"tokens_out":1988,"would_cite":false,"duration_ms":21357,"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":"A multi-view fake news detector that dynamically rewires a news hypergraph during convolution, and aligns propagation-tree and hypergraph embeddings by authenticity, reports the best accuracy and F1 among nine baselines on PolitiFact and…","keywords":["fake news detection","dynamic hypergraph","hypergraph structure learning","multi-view learning","contrastive learning","propagation tree","social media"],"falsifier":"Train DHy-MFND with the contrastive loss in Eq. (16) restricted to the training split, exactly as the loss must be if labels are not to leak; if the resulting PolitiFact or Gossipcop accuracy falls materially below the reported 92.81% or 98.84%, the reported superiority over baselines is not explained by the described method alone.","tokens_in":11972,"feed_emoji":"📰","tokens_out":8029,"duration_ms":70775,"temperature":0.7,"pith_summary":"This paper proposes a fake news detection model that combines three views of a news item: the text, its propagation tree on social media, and a hypergraph linking news pieces that share users, publication times, or entities. The key move is to treat the hypergraph structure as learnable: after each convolution layer the model rebuilds hyperedges from node and hyperedge embedding similarities, so the high-order relationships are optimized for the task rather than fixed in advance. It also adds supervised contrastive learning between the propagation and hypergraph views, pulling same-authenticity news together and pushing opposite-authenticity news apart. On PolitiFact and Gossipcop the model reports the best accuracy and F1 among nine baselines, with accuracy 92.81% and 98.84% respectively.","feed_headline":"Rewiring news hypergraphs on the fly beats nine fake news detectors","feed_subtitle":"Dynamic rewiring plus contrastive learning: 92.81% on PolitiFact, 98.84% on Gossipcop.","key_machinery":"The load-bearing mechanism is a dynamic hypergraph convolution block. At each layer, after nodes pass messages to hyperedges, a weighted cosine similarity matrix between news embeddings and hyperedge embeddings selects, per hyperedge, the top-$p_{thd}$ fraction of most similar nodes, producing a reconstructed incidence matrix $H_{re}$; a graph-level attention layer then fuses the original hypergraph with the reconstructed one and a text-derived hypergraph. A supervised InfoNCE contrastive loss between propagation-tree embeddings $X_{pro}$ and hypergraph embeddings $X_{hg}$ pulls same-label news together and pushes different-label news apart, making the learned embeddings authenticity-relevant.","core_discovery":"The paper claims that fake news detection improves when the relational structure used to compare news pieces is itself learned from the data, rather than fixed by handcrafted rules. DHy-MFND starts with a predefined hypergraph whose hyperedges group news by shared users, close publication times, or similar entities, then, during hypergraph convolution, recomputes these hyperedges from the current node and hyperedge embeddings using weighted cosine similarity, fuses the original and rebuilt structures with a graph-level attention layer, and repeats this rewiring at every layer. At the same time, a supervised InfoNCE loss aligns the propagation-tree embedding and hypergraph embedding of news with the same authenticity label and separates pairs with different labels. The fused multi-view embedding is classified by a softmax. The reported consequence is that this design beats nine baselines on both datasets, with the largest gains on PolitiFact, where hypergraph-based relational reasoning helps most.","pith_inferences":["Editorial inference: the dynamic structure-learning block is not news-specific; it should transfer to other hypergraph node-classification problems where initial hyperedges are noisy, such as rumor stance detection or document classification.","Editorial inference: the threshold $p_{thd}$ acts as a hard top-k selection, so replacing it with a differentiable soft assignment would let the structure learning be trained end-to-end with straight-through gradients, a direct variant the paper does not explore.","Editorial inference: the paper's runtime table shows Gossipcop takes 311.73 seconds per epoch versus 7.51 seconds for PolitiFact; scaling the dynamic rewiring to larger hypergraphs would require subsampling or incremental hyperedge updates, which the paper does not address."],"forward_implications":["The reported results imply that hand-built hyperedges are a floor, not a ceiling: letting the structure adapt during training improves both accuracy and F1 over the fixed-hypergraph baseline on both datasets.","Removing the propagation-tree view degrades the model more than removing the hypergraph view on Gossipcop, while the reverse holds on PolitiFact, so the three views carry complementary signal rather than redundancy.","Because the model needs only news text, propagation trees, and user interaction data, it can be deployed where large heterogeneous graphs with external knowledge bases are unavailable.","The ablation without contrastive learning drops about four accuracy points on PolitiFact, so the supervised alignment between the two views is a substantial contributor, not a marginal trick."],"supporting_citations":[{"why":"Supplies the predefined hypergraph construction strategy based on shared users, publication time, and similar entities, and the fixed-hypergraph baseline HGFND.","marker":"(Jeong et al. 2022)"},{"why":"Provides the UPFD dataset preprocessing, BERT-encoded news and user features, propagation-tree construction, and several propagation baselines.","marker":"(Dou et al. 2021)"},{"why":"Inspires the similarity-based hypergraph structure learning used in the dynamic rewiring module.","marker":"(Cai et al. 2022)"},{"why":"Supplies the InfoNCE contrastive loss that aligns propagation and hypergraph embeddings by authenticity label.","marker":"(Oord et al. 2019)"},{"why":"Motivates strengthening the source-news influence when constructing the text-derived hypergraph and provides a propagation-based baseline.","marker":"(Bian et al. 2020)"},{"why":"Provides the hypergraph neural network baseline HGNN and the two-stage node-hyperedge message-passing formulation used by the encoder.","marker":"(Feng et al. 2019)"},{"why":"Supplies the FakeNewsNet datasets from which the PolitiFact and Gossipcop benchmarks are drawn.","marker":"(Shu et al. 2020)"},{"why":"Gives the pre-trained language model used to encode news text into the initial node features for all views.","marker":"(Devlin et al. 2019)"}],"fun_headline_variants":["Rewiring hypergraphs on the fly improves fake news detection","Dynamic hypergraph learning: 92.8% on PolitiFact, 98.8% on Gossipcop","Multi-view model with dynamic hypergraphs and contrastive learning","Learn news hyperedges from data to spot fake news better","DHy-MFND: dynamic hypergraphs outsmart nine baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's headline results assume that validation and test labels are never used to build the training signal; the paper writes the contrastive loss over all news pieces using ground-truth labels and never states that the loss excludes the validation and test splits.","fun_headline_variants_meta":{"raw":{"variants":["Rewiring hypergraphs on the fly improves fake news detection","Dynamic hypergraph learning: 92.8% on PolitiFact, 98.8% on Gossipcop","Multi-view model with dynamic hypergraphs and contrastive learning","Learn news hyperedges from data to spot fake news better","DHy-MFND: dynamic hypergraphs outsmart nine baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1343,"prompt_tokens":959,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":575,"tokens_out":384,"duration_ms":4706,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:48:48.135021+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DHy-MFND with the contrastive loss in Eq. (16) restricted to the training split, exactly as the loss must be if labels are not to leak; if the resulting PolitiFact or Gossipcop accuracy falls materially below the reported 92.81% or 98.84%, the reported superiority over baselines is not explained by the described method alone.","supporting_citations":[],"review_version":1}