{"id":"c558c870-fdb6-436d-8c50-222e4f3fdaad","arxiv_id":"2412.01849","paper_version":2,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey taxonomizing directed graph neural networks into message-passing, eigenpolynomial, and sequence-based frameworks and re-reading them from a data-centric perspective.","lead":"This preprint is a survey of graph neural networks that work directly on directed graphs, organized around a three-part taxonomy: message-passing, spectral/eigenpolynomial, and sequence-based methods. It aims to be the first comprehensive review of directed GNNs from a data-centric viewpoint, useful for researchers mapping the field.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first comprehensive survey' claim rests on an unverified literature-priority assertion; no search protocol is reported and cited data-centric graph surveys may already cover directed GNNs, so the central novelty is not established.","rationale":"The reader identified the same load-bearing premise: the literature-priority claim in Section 1. My review of the full text finds no additional evidence that would change that assessment. The paper is an organizational review rather than a derivational contribution, so its primary testable claim is not a scientific result but a factual claim of absence. That claim is asserted with 'to the best of our knowledge' and without a search protocol, which makes it unverifiable from the manuscript alone. Given that the survey is otherwise a potentially useful synthesis, this does not warrant rejection; it warrants retaining the UNVERDICTED status until the priority claim is checked. I agree with the reader's framing and do not see a separate load-bearing concern that would move the verdict. The taxonomy's three categories are also asserted without formal disjointness criteria, but that is a secondary issue: even imperfect partitions can be useful, whereas the 'first survey' claim is binary and decisive for the paper's stated novelty.","tokens_in":31454,"tokens_out":3905,"duration_ms":37059,"concrete_test":"Retrieve the full texts of arXiv:2310.04987 ([138]) and arXiv:2309.10979 ([158]) and run a full-text search for 'directed' and 'digraph'. If either contains a dedicated section, taxonomy, or benchmark discussion specifically addressing directed GNNs, the priority claim is contradicted. Independently, run a structured bibliographic search on DBLP, arXiv, and Semantic Scholar with queries such as 'directed graph neural network survey', 'directed graph representation learning survey', 'digraph neural network review', and 'data-centric directed graph learning', restricted to works published before December 2024. If any retrieved survey has a scope comparable to Sections 1-4 of this paper, the 'first comprehensive survey' claim fails; if no such survey is found, the claim should be retained as supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in the Contribution list and the sentence 'to the best of our knowledge, no survey has yet specifically addressed directed GNNs' (Section 1), is that this is the first comprehensive survey of data-centric machine learning on directed graphs. The entire novelty depends on that absence, yet the manuscript provides no systematic evidence for it: there is no described search protocol, no database list (DBLP, arXiv, Google Scholar, Semantic Scholar), no query set, no inclusion/exclusion criteria, and no explicit comparison of scope against adjacent surveys. This is not merely a stylistic omission. The manuscript itself cites two data-centric graph learning surveys, [138] 'Data-centric graph learning: A survey' (arXiv:2310.04987) and [158] 'Towards data-centric graph machine learning: Review and outlook' (arXiv:2309.10979), both of which plausibly contain sections on directed graphs. If either survey, or any comparable prior survey, already organizes directed GNN methods under a data-centric or taxonomy-focused lens, the 'first comprehensive survey' contribution is void and the paper's main advertised novelty collapses. The three-way taxonomy may still be a useful organizational device, but it cannot by itself establish that the authors are filling a gap unless that gap is demonstrated. Since the absence claim is load-bearing and unverified, the survey's central contribution is currently unestablished rather than false.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a survey of directed graph neural networks (directed GNNs) framed from a data-centric perspective. It proposes a taxonomy with three frameworks—message-passing, eigenpolynomial-based, and sequence-based—and reviews representative directed GNN methods under each. It then re-examines these methods through three \"graph data understanding\" views (topological, spectral, sequence) and two improvement families (topological enhancement and node feature enhancement), catalogs applications across roughly a dozen domains, and closes with future research directions. The advertised contributions are: a claim to be the first comprehensive data-centric survey of directed GNNs, a novel taxonomy, a data-centric revisiting of existing methods, an industrial application overview, and a discussion of open challenges.","tokens_in":31665,"tokens_out":6271,"duration_ms":60109,"significance":"If the priority and coverage claims were established, this survey would be a useful organizational reference: it assembles a broad bibliography, gives compact formulas for spectral directed GNN methods, and connects directed GNN research to many application areas. The three-view data-centric framing is a reasonable pedagogical device, and the taxonomy, however imperfect, does give readers a structured entry point into a fragmented literature. The main weakness is that the survey's central novelty rests on an unverified \"first comprehensive survey\" claim, and the taxonomy's completeness is asserted rather than demonstrated. The paper is a competent descriptive survey, but it is not currently a definitive or fully auditable one.","major_comments":[{"comment":"The load-bearing novelty claim—\"to the best of our knowledge, no survey has yet specifically addressed directed GNNs\"—is not supported by any systematic evidence. The manuscript itself cites two data-centric graph learning surveys, [138] and [158], but does not state whether those surveys cover directed graphs, what their scopes are, or why they do not preempt the present contribution. No search protocol is reported: no databases, query set, date range, or inclusion/exclusion criteria. The authors should either provide a systematic scoping comparison against [138], [158], and other relevant graph-learning surveys, or soften the \"first comprehensive\" claim to a more specific and verifiable contribution.","section":"Section 1, Contributions bullet \"Comprehensive Review\""},{"comment":"The taxonomy is asserted to encompass three principal frameworks, but the criteria for assigning a method to exactly one framework are never stated. The placement of LightDiC [70] under the message-passing framework is an example of the ambiguity: the text describes it as a digraph convolution founded on the magnetic Laplacian, which has a spectral basis, yet it is classified under \"Propogator\" in Table 3. To make the taxonomy auditable, the authors should define the defining properties of each framework and give a rule or argument for exhaustiveness and mutual exclusivity.","section":"Section 3 and Table 3"},{"comment":"Eq. (2) presents the graph spectral filter as x∗gθ = U gθ(Λ) Uᵀ x = Σ θᵢ Tᵢ(L͂)x, which presupposes that the filter basis has an orthogonal eigenbasis. That assumption holds for real-symmetric or normal matrices, but not for general directed graph operators. The paper itself later acknowledges this issue in the HoloNet discussion, where it notes that directed filter bases lacking the real-symmetric property cannot be decomposed into complete orthogonal eigenvectors. Eq. (2) should therefore be qualified as applying only to diagonalizable/normal filter bases, and the conjugate transpose should be used where the basis is complex.","section":"Section 3.2, Eq. (2)"},{"comment":"The checkmark-based classification in Table 4 is not consistently supported by the text, and the meaning of each column is not defined. For example, DiRW [103] is described in Section 4.2.1 as a diffusion-based enhancement that transforms the adjacency matrix into an augmented matrix, but Table 4 shows no improvement checkmarks for DiRW. Similar discrepancies make the \"data-centric revisiting\" contribution hard to audit. The table needs a legend, explicit column definitions, and a method-by-method justification for each checkmark.","section":"Section 4.2.1 and Table 4"}],"minor_comments":[{"comment":"The definition of e_{ij} = (v_i, v_j) ∈ V is incorrect; the edge should be an element of E, not V. Also, \"Through out\" should be \"Throughout\".","section":"Section 2.1"},{"comment":"Several typos appear in the tables: \"Propogator\" should be \"Propagator\", and \"NeurlPS\" should be \"NeurIPS\".","section":"Table 3 and Table 4"},{"comment":"The text contains repeated typos: \"comlex-value\" should be \"complex-value\", \"eigenpolynimial\" should be \"eigenpolynomial\", and \"Markiv chain\" should be \"Markov chain\".","section":"Section 3.2"},{"comment":"Reference [126] is listed as \"Traffic-GGNN\" in Table 1 but as \"Traffic-GCNN\" in the text; the naming should be consistent.","section":"Section 5.1"},{"comment":"The method [49] is called \"EDGNN\" in the text and \"edGNN\" in Table 4; please use a single consistent name.","section":"Section 3.1 and Table 4"},{"comment":"The Framelet-MagNet paper is duplicated as references [76] and [77], and the text uses both numbers for the same work.","section":"References"},{"comment":"The sentence \"the necessitates for forecasting the directed perspectives of these interactions\" is ungrammatical and should be revised.","section":"Section 5.2"},{"comment":"The entry \"DawnGNNz\" appears to be a typo for \"DawnGNN\" (reference [31]).","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The editor should note that the \"first comprehensive survey\" claim is central to the paper's contribution and needs verification before acceptance. The manuscript cites two prior data-centric graph surveys that could plausibly cover directed graphs, and no systematic protocol is given for why they do not. Additionally, the survey's emphasis on the authors' own methods (LightDiC [70], ADPA [105], DiRW [103]) is noticeable; asking the authors to balance the coverage or justify the selection would strengthen the objectivity of the survey."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serviceable survey of directed GNNs, organized by a three-way taxonomy (message-passing, eigenpolynomial, sequence) and wrapped in a data-centric framing. The taxonomy is a reasonable way to sort the literature, and the applications table covers 11 domains with a lot of citations. If you need a starting point for directed GNN methods, this is a fine place to look.\n\nWhat's actually new: not much in the way of results — no new algorithm, theory, dataset, or benchmark. The contribution is organizational. The data-centric framing is mostly a re-description of existing method papers: graph construction and improvement are already implicit in models like DiGCN or MagNet. The paper does draw attention to directed graph quality as a design axis, which is a useful emphasis even if it's not a discovery.\n\nSoft spots: the biggest one is the literature-priority claim. Section 1 says 'to the best of our knowledge, no survey has yet specifically addressed directed GNNs' and the contributions call this 'the first comprehensive survey.' There's no search protocol, no list of databases or queries, no inclusion/exclusion criteria, and no explicit comparison against the two data-centric graph surveys they cite ([138] and [158]). If either of those already covers directed methods, the 'first' claim is wrong. That's a real gap in a survey whose novelty rests on being first. The taxonomy also isn't proven to be exhaustive — it's asserted. Minor issues: some model descriptions are thin (e.g., GNNDM is described in one sentence), and there's a mild self-citation pattern (LightDiC, ADPA, DiRW are the authors' own), though all three are legitimate works in the area.\n\nThat said, the survey's usefulness doesn't depend entirely on being first. The organization is clear, the background is accurate at the level I checked, and the future-directions section is sensible. I'd trust it as a starting point, not as a definitive authority on coverage.\n\nBottom line: send it to review. It's a competent survey that deserves referee time, but the authors should be pushed to either verify the 'first' claim with a documented search or soften it. For me, it's a cite-worthy reference for directed GNN methods, and I'd probably bring it to a reading group for the taxonomy discussion.","headline":"Useful survey of directed GNNs with a reasonable taxonomy, but the 'first comprehensive survey' claim is asserted without the search evidence to back it up.","tokens_in":32246,"tokens_out":2423,"would_cite":true,"duration_ms":21274,"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 claims to be the first dedicated data-centric review of directed graph learning and organizes directed GNNs into message-passing, eigenpolynomial-based, and sequence-based frameworks.","keywords":["directed graphs","graph neural networks","data-centric machine learning","survey","taxonomy","graph spectral theory","graph transformers","directed graph applications"],"falsifier":"Run a systematic literature search for a peer-reviewed survey published before this paper whose explicit subject is directed graph neural networks or directed graph representation learning; finding one with comparable scope would falsify the first-survey claim. To test the taxonomy, take a directed GNN method not listed in the paper and classify it: if its core mechanism cannot be assigned to message-passing, eigenpolynomial-based, or sequence-based without forcing, the claimed partition is not exhaustive.","tokens_in":31220,"feed_emoji":"🧭","tokens_out":6652,"duration_ms":58188,"temperature":0.7,"pith_summary":"This survey sets out to establish that directed graphs deserve their own data-centric treatment in graph machine learning, and that the existing literature can be organized into a single map. Its central claim is that it is the first dedicated survey of directed GNNs, and its main instrument is a three-part taxonomy: message-passing models that propagate along directed edges, eigenpolynomial-based spectral models that filter asymmetric graph Laplacians, and sequence-based models that tokenize graphs for attention. The paper argues that preserving edge directionality raises the representational ceiling of GNNs and that graph construction and improvement, not just model design, limit performance. A sympathetic reader would take away a structured guide to the field, a data-centric vocabulary for diagnosing model failures, and a catalog of applications across more than ten domains.","feed_headline":"First survey maps directed GNNs into three frameworks","feed_subtitle":"The first dedicated survey of directed GNNs, and a practical guide to when edge direction matters.","key_machinery":"The central organizational device is the survey's taxonomy of directed GNNs, defined by the machinery a model uses to respect edge directionality. Message-passing models split propagation and aggregation along incoming and outgoing edges; eigenpolynomial-based models replace the symmetric Laplacian with directed filter bases, such as the personalized PageRank transition matrix or the magnetic Laplacian, and approximate spectral filters with polynomials; sequence-based models tokenize graphs into node or subgraph sequences and apply attention with directed positional encodings. The accompanying data-centric pipeline (graph construction, graph improvement, and GNN learning) and the three graph views do the work of reframing each method as a statement about data quality rather than only architecture.","core_discovery":"The paper's core claim, stated on its own terms, is that no prior survey has specifically addressed directed GNNs, and that a data-centric review of this area is therefore both novel and needed. It proposes a taxonomy of directed GNNs into three principal frameworks: the message-passing framework, where propagation and aggregation operators are adapted to asymmetric edges; the eigenpolynomial-based framework, where graph filtering is done through filter bases and linear polynomials suited to non-symmetric and often complex-valued spectra; and the sequence-based framework, where graphs are tokenized into node or subgraph sequences and encoded with attention. The paper then revisits these methods from a data-centric perspective, reading each model through one of three graph views (topological, spectral, or sequential) and organizing data-improvement techniques into topological enhancement and node feature enhancement. Its further claim is that directed GNNs already serve more than ten application domains, and it closes with directions such as node feature denoising, scalability, robustness, and benchmarks for directed graphs.","pith_inferences":["A natural extension the paper leaves implicit is that the taxonomy could be applied to undirected GNNs by treating them as the special case where every edge is bidirectional, yielding a unified evaluation framework across directed and undirected graph learning.","If the data-centric premise is correct, benchmarks for directed graph learning should report graph-construction and augmentation choices alongside model architecture, since the survey implies that data quality bounds achievable performance.","The literature-priority claim is testable through a systematic citation and database search; even if an earlier directed-GNN survey appears, the three-framework taxonomy would remain a useful independent contribution."],"forward_implications":["If the survey's map is right, practitioners can choose a directed GNN family by the view of the data they trust: topology-based message passing, spectral filtering, or sequence-based attention.","Preserving edge directionality should be treated as a data-quality decision, not a model detail; degrading a directed graph to undirected form caps what any downstream architecture can learn.","The data-centric pipeline gives a shared vocabulary for failure analysis: a directed GNN underperforms either because the graph was poorly constructed or because the representation ignores directionality, not solely because the architecture is weak.","The catalog of applications implies that remodeling a domain as directed, for instance treating drug-drug interaction as an edge-level directed prediction task, can open modeling options that undirected formulations miss.","The identified gaps (node feature denoising, subgraph-token sequences, and scalable, robust directed GNNs) point to the concrete next targets for the field."],"supporting_citations":[{"why":"Prior survey on heterogeneous GNNs that the paper uses to show adjacent graph types already have dedicated reviews, establishing the gap for directed graphs.","marker":"[10]"},{"why":"Prior survey on hypergraph representation learning, one of the adjacent data-centric surveys the paper contrasts with the directed case.","marker":"[5]"},{"why":"Introduces the motif Laplacian for directed graphs and anchors the eigenpolynomial-based framework's filter-basis direction.","marker":"[88]"},{"why":"DiGCN uses personalized PageRank and $k$-order proximity in a directed spectral filter; a named representative of the eigenpolynomial-based framework.","marker":"[112]"},{"why":"MagNet introduces the magnetic Laplacian that later directed spectral and sequence models build on; central to the eigenpolynomial-based and positional-encoding discussion.","marker":"[152]"},{"why":"Dir-GNN formulates direction-aware propagation and aggregation for incoming and outgoing edges; a representative of the message-passing framework.","marker":"[95]"},{"why":"Directed Graph Transformers use dual encodings with separate query, key, and value matrices for source and target roles; anchors the sequence-based framework.","marker":"[123]"},{"why":"Prior data-centric graph learning survey that supplies the data-centric framing the paper carries over to directed graphs.","marker":"[138]"}],"fun_headline_variants":["Survey defines three frameworks for directed GNNs","Data-centric taxonomy for directed graph neural networks","Directed GNNs explained with three new lenses","A practical guide to when edge direction matters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that no earlier dedicated survey of directed GNNs exists, a claim asserted without a systematic search protocol; the taxonomy's exhaustive three-way partition is a second fragile premise.","fun_headline_variants_meta":{"raw":{"variants":["Survey defines three frameworks for directed GNNs","Data-centric taxonomy for directed graph neural networks","Directed GNNs explained with three new lenses","A practical guide to when edge direction matters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1634,"prompt_tokens":1020,"completion_tokens":614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":557}},"tokens_in":636,"tokens_out":614,"duration_ms":5978,"temperature":1.0,"reasoning_tokens":557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:42:43.598082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a systematic literature search for a peer-reviewed survey published before this paper whose explicit subject is directed graph neural networks or directed graph representation learning; finding one with comparable scope would falsify the first-survey claim. To test the taxonomy, take a directed GNN method not listed in the paper and classify it: if its core mechanism cannot be assigned to message-passing, eigenpolynomial-based, or sequence-based without forcing, the claimed partition is not exhaustive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MagNet introduces the magnetic Laplacian that later directed spectral and sequence models build on; central to the eigenpolynomial-based and positional-encoding discussion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Directed Graph Transformers use dual encodings with separate query, key, and value matrices for source and target roles; anchors the sequence-based framework."}],"review_version":1}