{"id":"77a25fb9-52a6-422f-9047-7578dc931dbe","arxiv_id":"2505.15834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using a new three-platform propagation dataset, a graph-based model that adapts to each platform's patterns outperforms single-platform and content-only detectors for fake news.","lead":"This paper introduces MPPFND, a dataset of fake news claims with user comments and propagation structure collected from YouTube, X, and Reddit, and reports that treating each platform's comment and reshare patterns separately improves automated fake news detection. A generalist might read it to understand why cross-platform social context is a promising signal for spotting misinformation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The all-platform F1 gain is not isolated from data quantity or claim overlap; Table 5 compares different claim subsets, so the evidence that cross-platform propagation differences improve detection is currently inconclusive.","rationale":"I focused on the experimental support for the central claim rather than the data-collection premise, because the headline-search issue, while important, is harder to falsify from the paper alone. The decisive question is whether the F1 gain in Table 5 is caused by cross-platform propagation differences. Table 5 does not hold the claim set or the amount of training data constant, and the paper's own interpretation admits that missing data drives the degradation. A pooled single-GNN control and a claim-level split would settle this. I do not dispute the value of the MPPFND resource or the descriptive platform analyses; the concern is confined to the detection experiment. This is a major revision but not a rejection, so I leave the conditional verdict unchanged.","tokens_in":11346,"tokens_out":9202,"duration_ms":96522,"concrete_test":"Restrict evaluation to the set of claims that have propagation on all three platforms, split by claim (all platforms of a claim in the same fold), and compare three models on this identical claim set: APSL on all platforms, the best single-platform APSL, and a pooled control that concatenates the three platform graphs into one GNN with the same parameter budget. If the all-platform APSL F1 exceeds the pooled control and the best single-platform model by more than run-to-run variance (e.g., three seeds), the cross-platform-difference claim is supported; otherwise it is an artifact of data volume or architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 5, where APSL on all platforms (F1 0.7045) beats single-platform X/Reddit/YouTube (0.6423/0.6545/0.6766). But these conditions are not matched: a claim is only included in a single-platform condition if it has posts on that platform, and the paper itself notes that degradation is largest 'when large amounts of propagation data are missing' (Section 'Effect of Modeling Propagation across Different Platforms'). The ordering of results tracks data volume, not platform-adaptive modeling: the conditions with more total propagation data score higher. There is no control that pools all platform graphs into a single GNN with the same capacity and training data, so the gain attributed to 'accounting for cross-platform propagation differences' could simply be more data and a more expressive per-platform architecture. In addition, the random 7:1:2 split is described over '4965 data samples' spanning three platforms; if the split unit is claim-platform pairs rather than claims, the same claim can occur in both training and test, inflating all reported numbers and further confounding the comparison. The central claim may be true, but this experiment does not yet establish it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents MPPFND, a fake-news detection dataset that links fact-checked claims from PolitiFact and Snopes to propagation graphs and user engagements on YouTube, X, and Reddit. The dataset contains 4,965 samples over 3,500+ claims and 440,000+ engagements. The authors analyze comment length, emotion, entity references, and graph statistics across platforms, reporting platform-specific and platform-shared patterns. They then propose APSL, which applies per-platform text adapters, per-platform GNN encoders, attention-based fusion guided by the claim text, and a platform-aware contrastive loss. Experiments compare APSL with content and propagation baselines on the new dataset, and ablations in Tables 4 and 5 are used to argue that multi-platform propagation and platform-adaptive modeling improve detection.","tokens_in":11607,"tokens_out":7074,"duration_ms":69099,"significance":"The dataset is a plausible contribution: it is among the first English multi-platform propagation datasets linking fact-checked claims to three major platforms, and the authors release data and code. The descriptive analyses of comment style, emotion, and entities are useful if the collection procedure is validated. The main claim, that accounting for cross-platform propagation differences improves fake-news detection, is not yet established because the supporting comparison in Table 5 is confounded and the principal results lack uncertainty quantification. If the matched experiments requested below confirm the effect, this would be a solid dataset-plus-method contribution; at present the central claim is defensible but unproven.","major_comments":[{"comment":"The claimed benefit of multi-platform propagation is not isolated from data quantity or claim coverage. The single-platform rows are evaluated only on claims that have posts on that platform, while the all-platform row uses the full claim set, and the paper itself notes that degradation is largest 'when large amounts of propagation data are missing.' Consequently the ordering in Table 5 may simply track the amount of training data per condition rather than the value of platform-adaptive modeling. I ask the authors to add a matched control that trains a single, non-platform-specific GNN on all propagation graphs pooled across platforms, with the same total data, parameter count, and training budget as APSL. In addition, report the per-condition number of claims, samples, and graph nodes, and rerun the comparison on the subset of claims that appear on all three platforms.","section":"Effect of Modeling Propagation across Different Platforms (Table 5)"},{"comment":"The random 7:1:2 split is described over '4965 data samples,' but the unit of splitting is not stated. If the unit is a claim-platform sample rather than a claim, the same claim can appear in both training and test through its posts on different platforms, inflating the reported numbers and obscuring the comparison. Please specify whether the split is by claim or by sample; if the split is by sample, provide a claim-level split version and report the number of claims in each partition.","section":"Experimental Setups"},{"comment":"The headline comparisons are averages of three runs with no variance or significance tests. For the BERT backbone, APSL F1 is 0.7045 versus UPFD-SAGE at 0.6821, a gap that may be real but is not assessable from three averaged runs. Please report mean and standard deviation (or individual runs) and run paired significance tests across repeated seeds for the key comparisons in Tables 3, 4, and 5.","section":"Main Results (Table 3)"},{"comment":"The collection procedure is underspecified: 'we create search queries based on the headlines to retrieve posts' does not state how queries were matched, how non-exact or duplicate posts were handled, or how retrieval quality was validated. Because every graph, statistic, and model input depends on the completeness and precision of headline search, please provide the matching rules, include retrieval statistics (for example, posts retrieved per claim and a manual verification sample), and discuss limitations due to API availability and platform search bias.","section":"Dataset Construction / Data Collection"}],"minor_comments":[{"comment":"The row for MPPFND contains the garbled entry 'X\\Youtube\\Reddit' and should be cleaned up to list the three platform names consistently.","section":"Table 1"},{"comment":"The typesetting of Equation (2) is ambiguous: the attention weight appears to include h_k^g both inside and outside the Softmax, and the role of the claim text c is not clear. Please define the query, key, and value explicitly and state the dimension d.","section":"Equation (2)"},{"comment":"The reference list contains entries that do not appear to be cited in the body, including Chalnick and Billman (1988), Feigenbaum (1963), Hill (1983), Matlock (2001), Newell and Simon (1972), Ohlsson and Langley (1985), and Shrager and Langley (1990); these should be removed or cited.","section":"References"},{"comment":"The text mentions 'gpt-4o-mini' as a baseline, but Table 3 lists 'GPT-4o'; please clarify which model was actually evaluated.","section":"Baselines and Tables"},{"comment":"The claim that Reddit exhibits a stronger echo chamber effect is based on comment-to-claim similarity, which is not a standard echo chamber measure; please justify this operationalization or use an established metric.","section":"Analysis of Comment Emotion and Echo Chambers"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely of interest to the fake-news detection community and the dataset release is valuable, but the central claim requires the matched experiments described in Major Comment 1 and a clear statement of the split unit. I would not reject on the current evidence, as the requested controls are feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper is worth reading for the dataset alone. MPPFND is the first English multi-platform propagation dataset for fake news I know of, pairing claims with propagation graphs from YouTube, X, and Reddit. The descriptive analyses of comment length, emotion, and echo-chamber effects are useful, and the authors make a plausible case that platform-specific signals exist.\n\nThe model APSL is a sensible combination of per-platform GNNs, attention fusion, and contrastive learning, and the ablations show each component contributes something. But the central claim—that accounting for cross-platform propagation differences improves detection—is not actually supported by the experiments as they stand. Table 5 compares the full multi-platform model against single-platform versions, but those conditions use different subsets of claims: a claim is only included in the single-platform condition if it has posts on that platform. So the comparison conflates platform-adaptive modeling with data quantity and claim coverage. The paper even notes that degradation is largest when large amounts of propagation data are missing. Without a control that pools all platform graphs into a single GNN with matched capacity and training data, the gain could simply reflect more data and a more expressive architecture.\n\nTwo more soft spots, both fixable. First, the 7:1:2 split is described over \"4965 data samples\" but never states whether the split unit is a claim or a claim-platform pair. With 3,500+ claims and 4,965 samples, that distinction matters for leakage. Second, there are no error bars or significance tests anywhere, so the modest F1 gap (0.7045 vs 0.6821) could be noise.\n\nNone of this kills the paper. The dataset is a real contribution, the analysis is thoughtful, and the authors are transparent about missing propagation data. The right fix is to add a properly matched control, clarify the split unit, and report variance. If they do that, the central claim will either be established or honestly qualified.\n\nThis is a paper for the fake news detection community, and specifically for anyone building cross-platform social context datasets. It deserves a serious referee—send it to review with the expectation of revision. I would cite the dataset even now.","headline":"A genuinely new multi-platform propagation dataset with useful platform analyses, but the central claim about cross-platform modeling is not yet supported by the experiments.","tokens_in":12148,"tokens_out":2731,"would_cite":true,"duration_ms":24681,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tracking a claim across YouTube, X, and Reddit improves automated fake news detection.","keywords":["fake news detection","multi-platform propagation","propagation graph","graph neural networks","platform-adaptive model","social media dataset","contrastive learning","MPPFND"],"falsifier":"Take a random sample of MPPFND claims, run independent searches on YouTube, X, and Reddit using both exact headlines and paraphrases, and compare the recovered posts with what the dataset contains; if a substantial share of engagements is missing or belongs to different stories, the propagation graphs and the reported APSL advantage rest on incomplete retrieval. A second check would rerun the detection comparison after adding the missed posts and see whether the multi-platform gain shrinks.","tokens_in":11144,"feed_emoji":"📡","tokens_out":13615,"duration_ms":114589,"temperature":0.7,"pith_summary":"This paper argues that fake news detection improves when propagation is tracked across several social platforms at once, rather than on one platform or from text alone. To support that claim, the authors build MPPFND, a dataset of more than 3,500 fact-checked claims with 440,000+ user engagements across YouTube, X, and Reddit, and show that engagement, comment length, sentiment, and echo-chamber behavior differ by platform. They then propose APSL, a graph-neural-network model that learns a separate propagation representation for each platform and combines them under the guidance of the claim text. On this benchmark, APSL reaches F1 0.7045, above the best baseline at 0.6821, and ablations show that removing any one platform or platform-specific component lowers performance.","feed_headline":"Cross-platform spread data lifts fake news detection to F1 0.7045","feed_subtitle":"A new dataset tracks the same claim across YouTube, X, and Reddit; platform-aware models beat single-platform baselines","key_machinery":"The load-bearing machinery is the Adaptive Propagation Structure Learning Network (APSL), applied to MPPFND's per-platform propagation graphs. APSL gives each platform its own graph neural network and a learnable platform vector $p_k$ that reweights comment embeddings, so structural and textual signals are not treated as if all platforms behave alike; a claim-guided attention module then fuses the platform representations, and a contrastive loss pulls together propagation features of same-label claims within each platform. The separate-GNN-plus-platform-adapter design is what lets the model turn cross-platform differences into a detection advantage.","core_discovery":"The central discovery is that the same news claim generates measurably different social context across platforms, and that exploiting those differences helps separate fake from true news. On MPPFND, fake news gets more engagement on YouTube and X than on Reddit, spreads across more platforms than true news, draws shorter comments on X and YouTube but not on Reddit, and evokes sentiment patterns that are informative on X and Reddit but nearly absent on YouTube. The authors build these observations into APSL, which trains a separate graph encoder per platform, adapts comment representations with platform-specific vectors, uses an attention mechanism to select propagation features relevant to the claim, and adds a platform-aware contrastive loss. This model reaches F1 0.7045, compared with 0.6821 for the strongest baseline, and ablations show that removing platform-specific components or any single platform degrades accuracy.","pith_inferences":["A testable extension not pursued in the paper: train APSL on two platforms and see whether its learned platform vectors let it adapt to a third platform without retraining, which would test how transferable platform adapters are.","Because the dataset was built by searching each platform with the claim headline, its coverage is bounded by search recall; benchmarking MPPFND against paraphrased-headline variants would show how much propagation is missed, especially on short-text platforms.","The observed imbalance, where fake claims spread across more platforms than true claims, could be turned into an explicit prior feature for detection rather than only a pattern learned implicitly by a graph encoder.","One could treat each platform's comment activity as a separate weak label source and use cross-platform agreement as a confidence score for early detection, before a full propagation tree has formed."],"forward_implications":["On MPPFND, a multi-platform propagation model beats content-only and single-platform models; APSL's F1 is 0.7045 versus 0.6821 for the best baseline and 0.6629 for a text-only content model.","Every platform contributes: removing any one platform's propagation lowers F1, and the largest drop comes from removing YouTube, whose propagation graphs dominate the dataset in node count.","Platforms are not interchangeable evidence sources: engagement separates fake and true claims on YouTube and X but not Reddit, while sentiment separates them on X and Reddit but barely on YouTube.","Fake claims tend to appear on more platforms than true claims, so multi-platform presence is itself a signal that detectors can exploit."],"supporting_citations":[{"why":"Supplies the label-mapping strategy that converts fact-checker labels into binary true/false, which MPPFND adopts.","marker":"(Nan et al., 2021)"},{"why":"FakeNewsNet is the single-platform propagation dataset that MPPFND extends by adding multiple platforms.","marker":"(Shu et al., 2020)"},{"why":"MCFEND is the multi-source benchmark MPPFND compares against in the dataset table.","marker":"(Li et al., 2024)"},{"why":"BiGCN is the propagation-graph baseline whose platform-pooled result motivates platform-adaptive modeling.","marker":"(Bian et al., 2020)"},{"why":"UPFD is the user-preference propagation baseline used in the main performance comparison.","marker":"(Dou et al., 2021)"},{"why":"UPSR handles incomplete propagation and is the strongest baseline APSL must beat.","marker":"(Wei et al., 2022)"},{"why":"BERT provides the text-encoder backbone for the main version of APSL and several baselines.","marker":"(Devlin et al., 2018)"},{"why":"RoBERTa provides the alternative text-encoder backbone for the second result block.","marker":"(Liu et al., 2019)"},{"why":"Contrastive learning objective that the platform-aware contrastive loss is modeled on.","marker":"(T. Chen et al., 2020)"}],"fun_headline_variants":["Cross-platform propagation differences improve fake news detection","New dataset tracks fake news spread across YouTube, X, Reddit","Platform-aware models outdo single-platform fake news detectors","Study: Fake news spreads differently on each platform—detection benefits","MPPFND dataset reveals cross-platform patterns in fake news"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset assumes that searching each platform by the claim's headline recovers the claim's real propagation structure; posts the search misses, or unrelated posts it pulls in, would distort every graph, platform analysis, and model trained on MPPFND.","fun_headline_variants_meta":{"raw":{"variants":["Cross-platform propagation differences improve fake news detection","New dataset tracks fake news spread across YouTube, X, Reddit","Platform-aware models outdo single-platform fake news detectors","Study: Fake news spreads differently on each platform—detection benefits","MPPFND dataset reveals cross-platform patterns in fake news"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3309,"prompt_tokens":839,"completion_tokens":2470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":2389}},"tokens_in":455,"tokens_out":2470,"duration_ms":17357,"temperature":1.0,"reasoning_tokens":2389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:56:12.080917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of MPPFND claims, run independent searches on YouTube, X, and Reddit using both exact headlines and paraphrases, and compare the recovered posts with what the dataset contains; if a substantial share of engagements is missing or belongs to different stories, the propagation graphs and the reported APSL advantage rest on incomplete retrieval. A second check would rerun the detection comparison after adding the missed posts and see whether the multi-platform gain shrinks.","supporting_citations":[{"cited_title":", Cao, J","cited_arxiv_id":null,"evidence_quote":"Supplies the label-mapping strategy that converts fact-checker labels into binary true/false, which MPPFND adopts."},{"cited_title":", Mahudeswaran, D","cited_arxiv_id":null,"evidence_quote":"FakeNewsNet is the single-platform propagation dataset that MPPFND extends by adding multiple platforms."},{"cited_title":", Xiao, X","cited_arxiv_id":null,"evidence_quote":"BiGCN is the propagation-graph baseline whose platform-pooled result motivates platform-adaptive modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UPSR handles incomplete propagation and is the strongest baseline APSL must beat."},{"cited_title":", Kornblith, S","cited_arxiv_id":null,"evidence_quote":"Contrastive learning objective that the platform-aware contrastive loss is modeled on."}],"review_version":1}