{"id":"cfe57132-0a88-4f42-b916-fbab12abf3b5","arxiv_id":"2501.15508","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"HML is a fake news detection framework that infers latent social networks with a Hawkes process and combines self-supervised multimodal learning with a graph transformer, claiming SOTA results on FakeSV and FVC.","lead":"This paper describes HML, a fake news detector that infers hidden social connections between news posts from their timing and content, without needing real retweet data. If it works, it could make fake-news detection feasible on platforms like TikTok and YouTube where cascades are hidden.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported SOTA is threatened by the transductive event-edge setup: a random 80:20 split still puts test news in the graph connected to labeled training news via event edges, so label propagation can inflate accuracy.","rationale":"Good-faith reading: the central claim is empirical—HML outperforms SOTA on FakeSV and FVC. The load-bearing condition is that the accuracy numbers measure generalization to unseen news rather than an artifact of the evaluation graph. The paper does not satisfy this condition: it reports a random 80:20 item split but builds the graph over the whole dataset before splitting, and the event edge type A1 connects each test item to labeled training items in the same event. Since fake/real status is highly event-correlated (e.g., a breaking event has many fake variants), the model can classify test items by their connectivity to labeled training items, not by the multimodal content strategy the paper claims. This concern is specific, testable, and independent of the internal consistency problems in the influence function (Eq. (2) vs Eq. (5)); even if those were fixed, the evaluation would still be ambiguous. No code, data, or error bars are provided, so the reader cannot independently verify the 89.14/91.02 numbers. A single re-run with an event-stratified or fully inductive split would settle whether the claimed SOTA is real. Therefore the reader's REJECT verdict is appropriate, and no change is needed.","tokens_in":12574,"tokens_out":5157,"duration_ms":52525,"concrete_test":"Run an event-stratified inductive evaluation: assign each event entirely to the training or test fold, build the graph from training news only, and then add test news without any edges to training nodes (or with all train-test edges masked). If HML's FakeSV accuracy drops by more than a small margin (say >2 points) from the reported 89.14%, the headline SOTA claim is inflated by transductive leakage through A1. As a control, also run a simple label-propagation classifier that uses only A1 edges; if it achieves accuracy comparable to HML, the event graph is carrying the classification signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Stage 1 (Social Latent Network Inference) builds a latent graph over all news nodes, with Edge Type A1 connecting news in the same event (Section 4.1, 'Heterogeneous Graph Construction'). The final adjacency in Eq. (11) is a learned combination of these edge types, and the Graph Transformer in Eq. (12) performs message passing over this full adjacency. The experimental setup section says only 'this paper divides the dataset 80:20 following the FakeSV benchmark'; it does not say the split is event-stratified, and no inductive protocol is described. Under a random item split, a test news item remains in the same event component as many training items. Because fake/real labels are strongly event-correlated, the event edges provide a path through which the model can effectively propagate label information from labeled nodes to test nodes, even though the focal loss in Eq. (13) is computed only on labeled nodes. The reported gains of +3.79 accuracy on FakeSV and +1.74 on FVC may therefore reflect transductive label propagation rather than the claimed multimodal content learning or latent network inference. This is not addressed by the ablation study, since every ablation uses the same random split and full graph.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HML, a multimodal fake-news detection method for short-video platforms where real retweet/cascade data are unavailable. Stage 1 infers a latent social network from event timings via an improved Hawkes process and constructs a heterogeneous graph whose edge types encode event membership, author attributes, location, counts, and content similarity. Stage 2 applies self-supervised unimodal augmentation and cross-modal contrastive learning to the text/image/audio/video features. Stage 3 combines edge types with a learnable attention mechanism, applies a graph transformer, and trains a classifier with focal loss. On FakeSV and FVC, the paper reports accuracy of 89.14% and 91.02%, respectively, claiming state-of-the-art results, and it also reports plugin improvements when the network-inference module is added to EANN, MCAN, and CAFE on Twitter and Weibo.","tokens_in":12906,"tokens_out":5446,"duration_ms":53550,"significance":"If the empirical claims hold, the paper addresses a genuine limitation of existing fake-news detection: many short-video platforms do not expose propagation cascades, so methods that rely on explicit social graphs are inapplicable. The idea of inferring a latent social graph from event timing and content attributes is timely, and the plugin experiments on Twitter and Weibo suggest the network-inference component may transfer to other settings. The paper also includes a reasonable ablation design separating Stage 1 and Stage 2 contributions. However, the central performance claim is currently undermined by a plausible transductive label-leakage path in the evaluation protocol and by the absence of any statistical uncertainty quantification. The methodological description of the Hawkes-process estimation is also incomplete. These are fixable in a revision, but they affect exactly the headline result.","major_comments":[{"comment":"The random 80:20 split described in the Experimental Setup is not event-stratified, and the graph used for message passing in Eq. (11)-(12) includes all news nodes, so test nodes remain connected to labeled training nodes through the Event edge type A1. Since fake/real status is strongly event-correlated, the Graph Transformer can propagate label information across this edge during training, inflating the reported Acc/F1 values. The paper does not describe an inductive protocol, a masked-edge evaluation, or an event-disjoint split; the ablation study uses the same split and therefore cannot rule out this leakage. Please re-run the evaluation under an event-disjoint or otherwise leakage-free protocol, or provide evidence that test nodes are isolated from labeled nodes.","section":"Experimental Setup and Section 4.1 (Heterogeneous Graph Construction, Eq. (11)-(12))"},{"comment":"The Experimental Setup promises \"interval estimation and K-S test\", but Table 1 reports only point estimates, and no confidence interval, standard deviation, or significance test appears anywhere. On FVC the F1 improvement over MMVD is +0.12%; without error bars this is indistinguishable from run-to-run noise and does not support the \"outperforms SOTA\" claim. Please report repeated-run statistics and significance tests for all headline numbers.","section":"Experimental Setup and Table 1"},{"comment":"In Eq. (5), prop(·) is defined as the Minkowski distance, i.e., a nonnegative, unbounded quantity, while the text preceding Eq. (2) states that prop(·) is a correlation in [-1,1]. The subsequent tanh(2·sim−1) mapping assumes a bounded similarity; using a raw distance makes the cross-event influence κ− (Eq. (2)) ill-defined and the values of γ uninterpretable. Please correct the definition or specify the normalization that turns a distance into a bounded similarity.","section":"Event-based Cascade Inference, Eq. (5)"},{"comment":"Section 4.1 states that the Hawkes-process parameters are estimated by the log-likelihood in Eq. (6), but the paper never gives the estimation procedure, update rules, or pseudocode; it also does not specify how the influence I(t) is converted into the adjacency matrix Ae. This makes Stage 1 non-reproducible and the claim that the graph is \"maximum likelihood\" unverifiable. Please add the full inference algorithm.","section":"Social Latent Network Inference (Stage 1)"}],"minor_comments":[{"comment":"The indices in Eq. (8) are inconsistent: the text lists ezi, ezj, ezp and then refers to augmented data ˜zj p, ˜zp p, but the equation does not define these quantities clearly. Please rewrite with consistent notation for original and augmented samples.","section":"Eq. (8) and Eq. (9)"},{"comment":"The noising process introduced in Eqs. (3)-(4) is not connected to the similarity calculation in Eq. (5); the paper should explain how this Gaussian noise enters prop(·) or remove the equations if they are not used.","section":"Event-based Cascade Inference, Eqs. (3)-(4)"},{"comment":"The phrase \"divides the dataset 80:20 following the FakeSV benchmark\" is ambiguous; please specify whether this is the official split of FakeSV and cite the exact protocol, since the split choice is central to the leakage concern.","section":"Experimental Setup"},{"comment":"The figure contains repeated labels such as ℒter1 that are not defined in the text; these should be corrected or removed for clarity.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The label-leakage concern in the random 80:20 split is serious enough that the current experiments, as reported, do not establish the SOTA claim. However, the issue is addressable with an event-disjoint or inductive evaluation protocol, and the other major issues (missing statistical tests, ill-defined similarity in Eq. (5), unspecified Hawkes estimation) are also within the scope of a major revision. If code or a detailed appendix is available, I would encourage the editor to request it, because the Stage 1 algorithm is currently not reproducible from the paper alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper attacks a real gap: fake news detection on short-video platforms where retweet cascades are unavailable. The combination of Hawkes-process latent network inference, self-supervised multimodal learning, and a graph transformer is new, and the plugin experiments on Twitter/Weibo suggest the latent network inference has some transferable value. The ablation study is reasonably thorough in showing each stage contributes. Credit where it is due: the problem framing is genuinely useful, and the overall architecture is more than just a bag of tricks.\n\nThe soft spots are real, and one is load-bearing. The random 80:20 split combined with event-based edges (Edge Type A1) puts test news in the same graph component as labeled training news. The graph transformer does message passing over the full adjacency, so labels can propagate through those event edges. The paper does not describe an event-stratified split or an inductive protocol, and every ablation uses the same setup. The reported gains (+3.79 accuracy on FakeSV, +1.74 on FVC) may therefore partly reflect transductive label propagation rather than the claimed content learning or latent inference. This is not a minor quibble; it undermines the empirical claim as stated.\n\nThere are also internal inconsistencies. The influence function defines prop(·) as tanh(2·sim−1) in Eq. (2), but Eq. (5) redefines prop(x_j) as a Minkowski distance, with no bridge between the two. The Gaussian noise equations (3)–(4) are introduced but their role in the model is never made clear. The paper mentions interval estimation and K-S test but reports no error bars or significance values. No code or data is provided.\n\nTo be fair, the core idea is salvageable. If the authors re-run with a leakage-free protocol and report error bars, the result could be meaningful. But as it stands, the SOTA claim is not trustworthy.\n\nThis paper is for researchers working on multimodal fake news detection, especially on short-video platforms. The problem setup and the plugin idea are worth knowing about, but the main empirical result needs to be re-established. I would send it to a serious referee because the idea is timely and the flaw is fixable, but only with the expectation of a major revision.","headline":"A timely idea with a useful problem framing, but the headline SOTA claim is not supported as written because the transductive graph setup can leak labels; worth engaging with, but only after the evaluation is fixed.","tokens_in":13372,"tokens_out":2842,"would_cite":false,"duration_ms":26588,"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":"Fake-news detector works without retweet cascades, hitting 89–91% accuracy on short-video benchmarks.","keywords":["fake news detection","multimodal","heterogeneous graph","latent network inference","Hawkes process","self-supervised learning","contrastive learning","short video platforms"],"falsifier":"Run HML under an inductive protocol where every test-set item is disconnected from the training graph by removing all edges incident to test nodes, or split per event so no event appears in both train and test. If accuracy stays near the reported 89–91%, the latent network generalizes; if it falls by several points, the gains come substantially from transductive label propagation through event edges.","tokens_in":12395,"feed_emoji":"🎬","tokens_out":5576,"duration_ms":49830,"temperature":0.7,"pith_summary":"Fake news on short-video platforms such as TikTok cannot be detected with the retweet-cascade graphs used for Twitter, because real cascades are largely hidden. HML sidesteps this by inferring a latent social network from news timestamps and social attributes: an improved Hawkes process estimates how much each news item influences others under the same event, and these influences are combined with attribute similarities into a heterogeneous graph. On top of that graph, self-supervised unimodal augmentation and cross-modal contrastive learning align text, image, audio, and video features before a graph transformer makes the final real/fake decision. On the FakeSV and FVC benchmarks the method reports 89.14% and 91.02% accuracy, beating previous best results by 0.12 to 4.39 percentage points, and the latent-network component alone improves existing detectors by up to 10 points when plugged into them.","feed_headline":"Fake-news detector hits 91% accuracy with no retweet data","feed_subtitle":"Inferring latent social networks from timestamps and attributes closes the gap left by hidden propagation on TikTok-style platforms.","key_machinery":"The load-bearing object is the inferred latent heterogeneous news graph $\\mathcal{H} = (V, E, O_E, A, X)$ in which every node is a news item and edges are typed by event, author introduction, news location, engagement statistics, and feature similarity. The event-edge submatrix $A_e$ comes from an improved Hawkes process $\\lambda^*(t) = \\lambda_0 + \\kappa(t)$, where $\\kappa^+$ sums self-excited past same-event influences $\\alpha e^{-\\beta(t-t_i)}$ and $\\kappa^-$ adds cross-event interference scaled by a correlation $\\tanh(2\\,\\mathrm{sim}(\\cdot,\\cdot)-1)$. This converts unobserved social influence into a prior adjacency matrix, which is then fused with attribute and similarity edges via learned attention weights $\\omega_{v\\leftrightarrow v}$ and passed through a graph transformer that masks attention with the adjacency, so structure shapes which news representations attend to one another before classification with focal loss.","core_discovery":"The central claim is that a latent cascade network—estimated instead of observed—can carry the social-context signal that retweet-based methods rely on. The authors model news influence intensity with an improved Hawkes process whose intensity is modulated by same-event self-excitation and cross-event interference, then threshold the pairwise influence matrix into event edges. Those edges join four other attribute-derived edge types in a heterogeneous graph where every node is a news item, not a user or venue. A self-supervised multimodal learning stage prunes and masks unimodal features, aligns them within and across modalities, and feeds the resulting representations together with the graph into a graph-transformer encoder trained with focal loss. The paper's evidence for the claim is the reported accuracy on FakeSV and FVC, plus plugin experiments showing consistent gains when the inferred network is added to three existing detectors.","pith_inferences":["Because event edges tie every news item in an event to the others, the same transductive wiring that may help the model also means the method's early-detection behavior is untested; evaluating with a split that hides all edges from future news would show whether the latent graph gives genuine lead time.","The Hawkes-process parameterization could be reused beyond fake news: any setting with timestamped items and weak or missing relational data, such as rumor cascades in messaging apps or coordinated posting detection, could adopt the same latent-influence matrix as a graph prior.","The plugin experiments suggest that the inferred network is a transferable representation, but the paper does not analyze what the learned edge-weight attention actually prefers; a weight-visualization study could turn the method into an interpretability tool for why a post is flagged fake.","A direct extension would be to replace the fixed Gaussian noising schedule with a learnable diffusion process, since the similarity computation used in $\\mathrm{prop}(\\cdot)$ already has a diffusion-like form; that could sharpen cross-event influence estimates on noisier platforms."],"forward_implications":["On FakeSV, HML reaches 89.14% accuracy and 89.22% F1; on FVC, 91.02% accuracy and 90.58% F1, beating all prior multimodal, unimodal, and LLM baselines.","Removing the latent network inference drops FakeSV accuracy by about 8 points, and removing the self-supervised multimodal learning drops it by about 10 points, so both stages are necessary for the reported result.","The latent-network component applied alone to EANN, MCAN, and CAFE improves accuracy by 0.31 to 9.86 points on Twitter and Weibo, indicating the inferred graph transfers to image-text fake news detection.","Because the method never needs observed retweet cascades, it is applicable to platforms where propagation data is unavailable or intentionally hidden.","The event-based Hawkes inference also captures cross-event interference, and ablations show that removing it costs 0.52% accuracy, evidence that connections between different events carry useful signal."],"supporting_citations":[{"why":"Supplies the FakeSV dataset, the primary benchmark with rich social context for short-video fake news, including event and attribute metadata.","marker":"Qi et al. 2023a"},{"why":"Supplies the FVC dataset of debunked and verified user-generated videos, the second benchmark used for evaluation.","marker":"Papadopoulou et al. 2019"},{"why":"Provides the Hawkes-process intensity model of information diffusion that the improved event-based cascade inference builds on.","marker":"Matsubara et al. 2017"},{"why":"Supplies the graph-transformer encoder used to aggregate the heterogeneous graph with masked attention and adjacency-aware weighting.","marker":"Zhu et al. 2024"},{"why":"Supplies the focal loss used as the classification loss to emphasize hard-to-classify news nodes.","marker":"Lin et al. 2017"},{"why":"Supplies the pruning-based augmentation strategy borrowed for unimodal content augmentation in the self-supervised learning stage.","marker":"Frankle and Carbin 2018"},{"why":"Provides the EANN baseline and dataset used to demonstrate that the latent-network component transfers to image-text fake news detection as a plugin.","marker":"Wang et al. 2018"}],"fun_headline_variants":["No retweet data? Infer the network and catch fake news","Latent network inference lifts fake-news detection without retweets","Fake news: infer hidden social ties to boost detection","91% accuracy, zero retweets: latent network inference does it","Retweet-blind fake news detection via latent network inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 80:20 random split still leaves test news connected to training news through event-type edges inside the same graph, and the paper does not show that this transductive wiring is label-free; if those edges leak class information during message passing, the reported accuracy overstates how well the method would perform on genuinely new events.","fun_headline_variants_meta":{"raw":{"variants":["No retweet data? Infer the network and catch fake news","Latent network inference lifts fake-news detection without retweets","Fake news: infer hidden social ties to boost detection","91% accuracy, zero retweets: latent network inference does it","Retweet-blind fake news detection via latent network inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3438,"prompt_tokens":895,"completion_tokens":2543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2459}},"tokens_in":511,"tokens_out":2543,"duration_ms":18710,"temperature":1.0,"reasoning_tokens":2459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:12:55.394036+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HML under an inductive protocol where every test-set item is disconnected from the training graph by removing all edges incident to test nodes, or split per event so no event appears in both train and test. If accuracy stays near the reported 89–91%, the latent network generalizes; if it falls by several points, the gains come substantially from transductive label propagation through event edges.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pruning-based augmentation strategy borrowed for unimodal content augmentation in the self-supervised learning stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FVC dataset of debunked and verified user-generated videos, the second benchmark used for evaluation."},{"cited_title":"A.; Li, L.; and Faloutsos, C","cited_arxiv_id":null,"evidence_quote":"Provides the Hawkes-process intensity model of information diffusion that the improved event-based cascade inference builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the graph-transformer encoder used to aggregate the heterogeneous graph with masked attention and adjacency-aware weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the focal loss used as the classification loss to emphasize hard-to-classify news nodes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EANN baseline and dataset used to demonstrate that the latent-network component transfers to image-text fake news detection as a plugin."}],"review_version":1}