{"id":"25a999f0-b4a3-4c95-bdc3-8f1ba1e83423","arxiv_id":"2506.18052","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that organizes propagation-based false information detection into homogeneous and heterogeneous categories, summarizing datasets, methods, and future directions.","lead":"This paper reviews how researchers detect false information on social networks by studying the way rumors spread, and sorts the methods into two groups: those using the reply chain of a single post, and those using wider social context. It is a reference survey that maps datasets, methods, and open problems for newcomers to the field.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'comprehensive' claim is unsupported: no literature selection protocol is given, and repeated misattributions of GAT (cited to Vieweg et al. 2010 and to Liu et al. 2018) show the method-to-reference mapping is unreliable.","rationale":"The reader's CONDITIONAL verdict is appropriate. The taxonomy is a plausible organizing device, and the method summaries broadly align with the literature I can check. The issue is not internal inconsistency in the taxonomy but the lack of verifiable support for the 'comprehensive' label and observable citation errors. These are fixable in revision: add a methodology subsection documenting the search and screening process, and correct or verify all references, especially GAT. This does not require rejecting the paper; it requires conditioning acceptance on those corrections. I therefore recommend UNCHANGED relative to the reader's CONDITIONAL verdict. I partially agree with the reader's weakest assumption: selection without a protocol is the main load-bearing gap, and the citation errors are concrete evidence that the same unverified mapping also contains inaccuracies. My proposed test would settle whether the omissions and misclassifications are material.","tokens_in":17545,"tokens_out":6752,"duration_ms":80112,"concrete_test":"Run a systematic literature search with explicit inclusion criteria over DBLP, Scopus, and arXiv for 2020-2025 using queries such as 'propagation tree rumor detection', 'homogeneous graph fake news', 'heterogeneous social context fake news', and 'LLM rumor detection'. Compare the retrieved set of relevant propagation-based methods with the methods reviewed in Sections 4.3 and 5.3. Report (a) the number of unique relevant methods omitted, and (b) the number of reviewed methods whose category or cited technique contradicts the original paper, starting with DUCK's use of GAT and SePro's use of GAT. If omitted relevant methods exceed 10% of the reviewed set, or if any method's category or citation is wrong, the 'comprehensive' claim is weakened and requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the homogeneous/heterogeneous taxonomy gives a comprehensive and accurate map of propagation-based false information detection. That claim depends on two unverified premises: (1) the reviewed method set is representative, and (2) each method's category and summary faithfully match the cited original. The paper gives no evidence for (1): there is no search protocol, no inclusion/exclusion criteria, and no completeness check, so the reader cannot tell whether omitted propagation-based work would break the taxonomy. There is direct evidence against (2): in Section 5.3.1, DUCK is described as using 'Graph Attention Networks (GAT) Vieweg et al. [2010]', but Vieweg et al. 2010 is a crisis-communication study, not the GAT paper; in Section 4.3.3, SePro is said to 'employ Graph Attention Networks Liu et al. [2018]', and Liu et al. 2018 is an attention-based misinformation paper, not GAT. GAT is a load-bearing component of several reviewed methods, so these misattributions suggest the method-to-reference mapping was not verified. A survey's value is exactly this mapping; if its boundaries are undocumented and its labels demonstrably unreliable, the 'comprehensive review' claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of false information detection methods that exploit information propagation on social networks. It organizes the field with a new taxonomy: homogeneous propagation-based methods, which use single-type propagation graphs formed by comments and retweets of a source post, and heterogeneous propagation-based methods, which additionally incorporate user metadata, semantic information, and cross-platform discussion. For each category the paper provides a formal problem formulation, descriptions of commonly used datasets, and a review of representative methods, followed by future research directions such as unified benchmarks, diverse information carriers, and mitigation strategies.","tokens_in":17710,"tokens_out":4274,"duration_ms":50366,"significance":"If the taxonomy and coverage were fully accurate and comprehensive, this survey would be a useful reference for researchers entering the area, particularly because it covers propagation-based methods beyond GNN-only surveys, including robustness, contrastive learning, and LLM-augmented detection. The formal problem formulations in Sections 4.1 and 5.1 and the dataset statistics in Tables 1 and 2 are clear strengths. However, the survey's value as a reference depends on the reliability of its method-to-citation mapping and on the completeness of its coverage; both are currently in question, which limits the significance of the contribution until those issues are addressed.","major_comments":[{"comment":"The manuscript repeatedly claims to provide a 'comprehensive review' and lists 'Comprehensive Review and Summary' as a core contribution, but it does not describe any systematic literature selection methodology. There is no search protocol, inclusion/exclusion criteria, publication timeframe, database list, or completeness check. Because the main contribution of a survey is precisely its coverage and mapping of the field, the absence of this methodology means a reader cannot verify whether omitted propagation-based methods would alter the taxonomy or its subcategories. Please add a methodology subsection that documents how papers were retrieved, screened, and selected, and acknowledge residual coverage limitations.","section":"Abstract and Section 1 (Contributions)"},{"comment":"The citation for Graph Attention Networks is misattributed in two places. In Section 5.3.1, DUCK is said to use 'Graph Attention Networks (GAT) Vieweg et al. [2010]', but Vieweg et al. [2010] is a crisis-communication study on microblogging during natural hazards, not the GAT paper. In Section 4.3.3, SePro is said to 'employ Graph Attention Networks Liu et al. [2018]', and Liu et al. [2018] is an attention-based misinformation identification paper, not the GAT architecture. Because GAT is cited as a load-bearing component of several reviewed methods, these errors show that the method-to-reference mapping has not been systematically verified, which is a central quality criterion for a survey. Please correct these citations and audit the rest of the reference list for similar mismatches.","section":"Section 5.3.1 (DUCK) and Section 4.3.3 (SePro)"},{"comment":"The taxonomy's second-level split for heterogeneous methods is unbalanced: 'Other Context' is a residual category that lumps together temporal information, domain information, keyword/semantic relations, and cross-platform discussion. As presented, this subcategory does not offer the analytic contrast promised by the taxonomy, because it is defined negatively as whatever is not user-related. Consider either refining this subcategory into more specific types (e.g., temporal, semantic, cross-platform) or re-framing the taxonomy as a three-way or four-way division, so that the claimed systematic organization is not just a two-bin split.","section":"Section 5.3.2 (Other Context)"}],"minor_comments":[{"comment":"There are typographical errors in author names; for example, 'V osoughi' and 'V ousoughi' should be 'Vosoughi' (reference for Vosoughi et al. [2018]).","section":"References"},{"comment":"The sentence 'RvNN and BiGCN utilize recursive neural models Cho et al. [2014] and graph neural networks Kipf and Welling [2016]' is ambiguous; it should state explicitly that RvNN uses recursive neural networks and BiGCN uses graph convolutional networks, with citations placed accordingly.","section":"Section 4.3.1"},{"comment":"The dataset descriptions state that PHEME5 and PHEME9 are similar in construction, but the table lists differing counts for comment/retweet posts (103,212 for PHEME5 versus 105,354 for PHEME9); it would help readers to explain whether the numbers reflect the original releases or post-processing at the time of writing.","section":"Section 4.2 and Table 1"},{"comment":"PSIN is described as using 'three variants of Graph Attention Networks' but the citation for GAT is not provided; the reference list already lacks the canonical GAT paper (Veličković et al., 2018), which should be added to avoid further misattribution.","section":"Section 5.3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a CS/SI venue and the taxonomy is potentially useful, but the two central problems—undocumented coverage and citation mismatches—need to be fixed before publication. The authors should be asked to provide a detailed response explaining how the method-to-citation mapping will be audited, since a survey with demonstrable reference errors cannot be accepted as-is. I also note that Zhang et al. [2025] appears both as a reviewed method (KPG in Section 4.3.2) and as the source of the Weibo22 dataset (Section 4.2), which is not improper but should be clearly flagged to avoid the appearance of over-reliance on the authors' own work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serviceable survey with a genuinely useful organizing idea — splitting propagation-based detection into homogeneous vs. heterogeneous — but the 'comprehensive' label is not earned, and the citation mapping has real reliability problems. I'd send it to review, but I'd require the authors to audit their references and document how they chose the papers.\n\nWhat's actually new: the homogeneous/heterogeneous framing is not just repackaging. It tracks a real methodological choice in the literature: some methods build a tree of post–comment interactions, others fold in users, cross-platform discussion, and other context. The formal problem formulations in Sections 4.1 and 5.1 are clean and genuinely help a newcomer see what task each family is solving. The dataset summaries are mostly accurate and the statistics tables are handy. The future-work section on a unified benchmark is sensible.\n\nWhere it's soft: First, the authors call this comprehensive but never say how they found the papers. There's no search protocol, no inclusion/exclusion criteria, no completeness check. For a survey whose value is organizational, that's not a minor omission — it makes the coverage claim unverifiable. Second, the method-to-reference mapping is demonstrably unreliable. DUCK is said to use GAT 'Vieweg et al. [2010]' — that's a crisis communication paper, not GAT — and SePro is said to use GAT 'Liu et al. [2018]', which is an attention-based misinformation paper, not the GAT paper. These are load-bearing components of the described methods, and if two are wrong, a reader can't trust the rest. The paper also cites the authors' own KPG paper (Zhang et al. 2025) as both a method and the Weibo22 dataset source; that's fine, but worth a sanity check.\n\nThe taxonomy itself still holds up. The organization is clear, the narrative is readable, and the problem definitions are a real service. I just wouldn't call it comprehensive without the missing methodology, and the citation errors need fixing before publication.\n\nWho's this for: a graduate student or researcher new to rumor detection who wants a map of propagation-based methods. It will be useful after revision. For an editor: yes, send to peer review; the framing is worth refereeing, but make sure a reviewer with deep knowledge of the area checks the citations.\n\nRecommendation: engage with it, require revision, don't desk-reject.","headline":"A useful homogeneous/heterogeneous organizing frame for propagation-based rumor detection, but the 'comprehensive' claim is unverified and the method-to-reference mapping has real reliability problems.","tokens_in":18255,"tokens_out":2334,"would_cite":true,"duration_ms":23087,"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 survey organizes false-information detection by propagation scope, separating methods that use the comment-retweet tree from methods that use the full social-context graph, and maps datasets and gaps along that split.","keywords":["false information detection","rumor detection","fake news","propagation on social networks","homogeneous propagation","heterogeneous propagation","graph neural networks","survey taxonomy"],"falsifier":"Identify the set of propagation-based detection papers published before June 2025 and code each against the homogeneous/heterogeneous dichotomy; if a sizable cluster uses both propagation scopes simultaneously or uses propagation signals with no graph structure at all, then the proposed taxonomy does not organize the whole field as claimed.","tokens_in":17242,"feed_emoji":"🕸️","tokens_out":8159,"duration_ms":86627,"temperature":0.7,"pith_summary":"False information on social media leaves traces in how it spreads, and this survey tries to establish that those traces are a usable organizing principle for the whole detection literature. The paper's central move is a taxonomy that splits propagation-based methods into homogeneous propagation, where a source post's comments and retweets form a single tree-shaped graph, and heterogeneous propagation, where the graph also includes users, friendships, keywords, and cross-platform discussion. Around this dichotomy the survey assembles formal problem definitions, dataset statistics, and method summaries for each branch. A reader gets both a map of existing work and a set of concrete gaps: no unified benchmark, text-dominated carriers, and little machinery for reducing harm after detection. The value of the paper, if its coverage holds, is that researchers can locate any method by what propagation evidence it uses and see what neighboring approaches already exist.","feed_headline":"Survey sorts fake-news detection by how it spreads","feed_subtitle":"New taxonomy splits methods by comment-tree versus social-context evidence, mapping datasets and gaps.","key_machinery":"The load-bearing distinction is between two graph abstractions. A homogeneous propagation graph $\\mathcal{G}_p=(V_p,E_p,X_p)$ is a tree whose nodes are the source post and its comments or retweets, whose edges are reply and retweet relations, and whose node features are post texts; with it, rumor detection becomes graph-level classification of the tree root. A heterogeneous social context graph $G=(V,E)$ has multiple node types---news posts, comment/retweet posts, and users---and multiple edge types such as publishing, commenting, and following, which turns detection into node classification over a richer graph. The survey uses these two formalizations to state the task each method solves, then organizes methods within each branch: homogeneous methods are grouped by propagation dynamics, robustness, and LLM augmentation, while heterogeneous methods are grouped by user-related versus other context. The taxonomy itself is the machinery that lets otherwise scattered papers be compared by the evidence they consume.","core_discovery":"The paper's central claim is that propagation-based false-information detection should be divided by the scope of propagation modeled. Homogeneous propagation methods use only the source post together with its comments and retweets, represented as a tree-structured graph where every node is a post; the task is graph-level veracity classification. Heterogeneous propagation methods build a graph with multiple node types---posts, news articles, users, and cross-platform discussions---and treat veracity prediction as node classification over that richer context. The survey further splits homogeneous methods into those modeling propagation dynamics, those improving robustness against noisy or adversarial comments, and those using large language models to interpret propagation; heterogeneous methods are split into user-related context and other context. The authors argue that this taxonomy captures the varying scope and complexity of information dissemination and that previous surveys, focused mainly on deep text models or on graph neural networks, missed this organizing viewpoint.","pith_inferences":["A consequence the authors do not develop is that their dichotomy suggests a missing middle: methods that start from the homogeneous tree and then selectively add heterogeneous user nodes, which could offer a cost-controlled way to scale evidence from single-platform threads to full social context.","If propagation scope is the organizing variable, then detection performance should degrade in a measurable way as the observation window shrinks; the early-detection and diffusion-prediction methods surveyed could be compared on a single curve of accuracy versus fraction of propagation observed, which the proposed benchmark could provide.","The corrective-bot proposal implicitly shifts the success metric from detecting false posts to reducing their spread; testing that would require new datasets with correction posts, user exposure, and subsequent sharing behavior, a gap the survey itself notes."],"forward_implications":["Each homogeneous-propagation method takes the same formal object, an event graph $\\mathcal{G}_p=(V_p,E_p,X_p)$ rooted at the source post, so shared datasets like PHEME and Weibo support direct method comparison within this branch.","Because robustness work treats real propagation as noisy, contrastive learning, adversarial training, and uncertainty-weighted edges are recurring tools; evaluation should therefore include noise or perturbation sensitivity, not just average accuracy.","LLM-augmented methods position large language models as readers of propagation context rather than replacements for graph structure, which points to hybrid architectures where the LLM supplies rationales and the graph supplies topology.","Heterogeneous methods demonstrate that user profiles, friendship relations, and cross-platform mentions carry veracity signal beyond the comment thread; models that ignore these inputs are working with strictly less evidence.","Future benchmark work as framed by the authors would replace arbitrary train/test splits with topic- or time-based splits and shared evaluation scripts, making reported numbers comparable and testing generalization rather than memorization."],"supporting_citations":[{"why":"Supplies the false-information taxonomy (misinformation versus disinformation, opinion-based versus fact-based) that grounds the survey's definitions.","marker":"Kumar and Shah [2018]"},{"why":"Provides the standard definition of fake news and the data-mining perspective that the survey positions itself against.","marker":"Shu et al. [2017]"},{"why":"Establishes the homogeneous propagation paradigm by modeling comment and retweet trees with recursive neural networks.","marker":"Ma et al. [2018]"},{"why":"Introduces the bi-directional propagation model and the event-as-propagation-tree formulation used throughout Section 4.","marker":"Bian et al. [2020]"},{"why":"Supplies the Weibo rumor dataset and the recurrent-network baseline that many homogeneous methods build on.","marker":"Ma et al. [2016]"},{"why":"Provides the PHEME datasets and the conversational-thread analysis that define standard homogeneous propagation benchmarks.","marker":"Zubiaga et al. [2016b]"},{"why":"Represents the user-preference branch of heterogeneous methods that the survey groups under user-related context.","marker":"Dou et al. [2021]"},{"why":"Supplies the FakeNewsNet-based datasets with user profiles, friendships, and cross-platform discussion used by heterogeneous methods.","marker":"Shu et al. [2020]"},{"why":"Provides the Weibo22 dataset and the key-propagation-graph generation method used as a current example and future-work anchor.","marker":"Zhang et al. [2025]"}],"fun_headline_variants":["New taxonomy sorts fake news by propagation scope","Survey divides rumor detection by comment trees vs context","Propagation-based detection: tree vs multi-node taxonomy","Fake news survey keys on how info spreads on networks","Homogeneous vs heterogeneous: new spread-aware taxonomy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's central claim presupposes that the methods it manually selected are representative of propagation-based detection research, since no systematic search protocol, inclusion criteria, or completeness check is described.","fun_headline_variants_meta":{"raw":{"variants":["New taxonomy sorts fake news by propagation scope","Survey divides rumor detection by comment trees vs context","Propagation-based detection: tree vs multi-node taxonomy","Fake news survey keys on how info spreads on networks","Homogeneous vs heterogeneous: new spread-aware taxonomy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1388,"prompt_tokens":867,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":483,"tokens_out":521,"duration_ms":5925,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:22:58.234863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Identify the set of propagation-based detection papers published before June 2025 and code each against the homogeneous/heterogeneous dichotomy; if a sizable cluster uses both propagation scopes simultaneously or uses propagation signals with no graph structure at all, then the proposed taxonomy does not organize the whole field as claimed.","supporting_citations":[],"review_version":1}