{"id":"39cabe19-ef84-4147-84d8-b70da5cf1a44","arxiv_id":"2502.06927","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NOL-GAT lets every node select its own k-hop neighborhood per layer, and the paper reports strong fake news detection gains under low label rates.","lead":"This paper proposes NOL-GAT, a graph neural network that lets each news article in a similarity graph learn which range of neighbors (hop order) to draw information from at each layer. The authors report large accuracy and F1 gains over six baselines across five fake news datasets, especially when only 10 to 30 percent of labels are available.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The large reported accuracy gains rest on an unverified adaptation of TextGCN/LSTM-CP-GCN and an unreported per-dataset K choice; until a clean re-run is provided, the central outperformance claim is not established.","rationale":"The reader's weakest assumption — that the baseline adaptation in Section 6.3 is unfair and the closest related methods (DHGAT, LOSS-GAT) are omitted — is exactly the load-bearing weakness I identify. If the baselines are weakened, the central empirical claim collapses regardless of the soundness of the proposed mechanism. I also note an additional concrete gap: the per-dataset KNN parameter k and hop-network order φ are not reported for the main results, even though k is tuned in Section 6.4.3, so the comparison may involve favorable test-set selection for NOL-GAT. This reinforces the reader's CONDITIONAL verdict rather than changing it: the authors should release verified code, report all hyperparameters per dataset, and re-run the comparison with faithful baseline implementations. I do not see an internal inconsistency in the model formulation itself that would independently refute the approach; the issue is purely empirical verifiability. Hence I recommend no change to the reader's verdict.","tokens_in":20452,"tokens_out":6075,"duration_ms":58865,"concrete_test":"Run the released NOL-GAT code and the original TextGCN implementation (word–document graph) and original LSTM-CP-GCN implementation on the same five datasets, same labeled splits, and a single fixed k (e.g., k=5) for all methods, and report the per-dataset k and hop-network φ in the tables. If NOL-GAT's accuracy/F1 margin over TextGCN or LSTM-CP-GCN shrinks to below ~5 points on any dataset, the headline outperformance claim is not supported by the current protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NOL-GAT significantly outperforms six baselines, especially at 10% labels. The credibility of this claim depends entirely on the fairness and completeness of the experimental protocol in Section 6.3. Two load-bearing problems are visible. First, TextGCN and LSTM-CP-GCN are described as 'following the original model architecture' while 'a two-layer GCN with the same number of hidden units is used' and all methods share Doc2Vec features and hyperparameters. TextGCN's mechanism is specifically the word-document heterogeneous graph; LSTM-CP-GCN's mechanism is the sentence graph with CP-decomposed co-occurrence weights. Reducing both to a generic two-layer GCN on a news-only KNN graph removes the very components that define them, so these baselines are likely artificially weak. Second, Section 6.4.3 tunes the KNN graph parameter k and finds the optimum at 6–7, but the main results in Tables 3–7 never report which k was used for NOL-GAT or for the baselines, and the hop-network neighborhood order φ (Algorithm 1 input k) is also undisclosed. If the reported numbers were selected from the best k for NOL-GAT while baselines used a fixed default, the margins (e.g., +0.19 accuracy over TextGCN on Fake.Br at 10% labels) would be inflated by test-set tuning. Neither issue invalidates the architecture, but together they mean the strongest empirical claim is not yet verifiable from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NOL-GAT, a graph attention architecture for semi-supervised fake news detection in which each node, at each layer, learns a categorical distribution over neighborhood orders (k-hop distances) using a Gumbel-Softmax estimator. A hop network (Φ) predicts the order, and an embedding network (Ψ) aggregates neighbors at the selected order. The architecture is evaluated on five datasets (Fake.Br, Fact-checked News, FakeNewsNet, FakeNewsDetection, FakeNewsData) at 10%, 20%, and 30% labeled-data rates, against six baselines (CO-GNN, L2Q, TGNCL, LSTM-CP-GCN, TextGCN, GATv2). The paper reports ten-run means and standard deviations for accuracy and macro-F1, and claims significant gains, especially in low-label settings, along with qualitative advantages such as mitigating over-squashing and reducing computational complexity.","tokens_in":20756,"tokens_out":3695,"duration_ms":33292,"significance":"If the reported results are taken at face value, the contribution is a simple and plausible mechanism: replacing fixed-depth message passing with per-node, per-layer hop selection, which is differentiable via Gumbel-Softmax and can be implemented as a small modification of GATv2. The GATv2 ablation in Figure 3 supports the claim that the adaptive hop-selection mechanism, rather than merely added capacity, drives the improvement. The manuscript also provides public code and reports ten-run means and standard deviations, which is good experimental practice. However, the significance is currently contingent on the fairness and completeness of the experimental protocol: the adaptation of TextGCN and LSTM-CP-GCN to a generic two-layer GCN is not described in sufficient detail, the KNN graph degree is tuned but not reported in the main tables, and the closest related methods from the same group (DHGAT, LOSS-GAT) are omitted. Until these issues are resolved, the headline outperformance claim is not verifiable from the manuscript.","major_comments":[{"comment":"The baseline configurations for TextGCN and LSTM-CP-GCN are not faithful to their original architectures. The text states that 'For TextGCN and LSTM-CP-GCN, a two-layer GCN with the same number of hidden units is used, following the original model architecture.' TextGCN's original mechanism is a heterogeneous graph with word and document nodes plus pretrained word embeddings, and LSTM-CP-GCN uses a sentence graph with CP-decomposed co-occurrence weights and LSTM-generated features. Reducing both to a plain two-layer GCN on the Doc2Vec-based KNN graph removes precisely the components that define these methods, which is likely to weaken them substantially. Please specify the exact adapted architecture and feature inputs for each baseline, or run the original implementations with their native graph constructions, and state whether any hyperparameters were re-tuned for the adapted versions.","section":"Section 6.3"},{"comment":"The main results do not report the KNN graph degree k used for each method. Section 6.4.3 tunes k and reports that the optimum is generally 6–7, but Tables 3–7 do not state whether NOL-GAT and each baseline used the same k, the per-dataset optimal k for NOL-GAT, or a fixed default for the baselines. If the reported NOL-GAT numbers were selected from the best k per dataset while baselines used a single un-tuned k, the reported margins (e.g., +0.19 accuracy over TextGCN on Fake.Br at 10% labels) would partly reflect test-set selection rather than the architecture. Please report the exact k for every method and every dataset, and, if k was tuned for NOL-GAT, provide a sensitivity analysis for the baselines over the same k values.","section":"Section 6.4.3 and Tables 3–7"},{"comment":"The closest related methods, DHGAT [46] and LOSS-GAT [28], are discussed in the related work but are absent from the experimental comparison. Both are semi-supervised GAT-based fake news detectors by the same research group; DHGAT in particular uses the same decision/representation network split as NOL-GAT, differing mainly in the action space (neighborhood type vs. hop order). Because the contribution is framed as an improvement in adaptive neighborhood selection, the absence of these methods from Tables 3–7 makes it impossible to assess whether NOL-GAT offers a genuine advance over the authors' own prior work. Please add these comparisons or justify their omission with concrete reasons.","section":"Section 2 and Section 6.2"},{"comment":"The claim that NOL-GAT 'significantly outperforms' baselines is not supported by any significance test. The tables report means and standard deviations over ten runs, but no paired t-tests, confidence intervals, or effect-size statistics are provided. While many margins appear large, some comparisons have overlapping variation (e.g., Table 4, 30% label case for Co-GNN vs. TextGCN), and no formal test is given. Please add appropriate statistical tests for the headline comparisons, or qualify the 'significant' language accordingly.","section":"Abstract and Section 6.4"}],"minor_comments":[{"comment":"The baseline named 'L2Q' is cited as [43], but reference [43] is titled 'Learning How to Propagate Messages in Graph Neural Networks' (L2P). Please clarify whether the baseline is L2P, a renamed variant, or a different method, and ensure the citation and name are consistent.","section":"Section 6.2"},{"comment":"The claims about mitigating over-squashing, improving information flow, and reducing computational complexity are stated as advantages but are not measured or quantified anywhere in the experiments. Please either provide supporting measurements (e.g., over-squashing metrics, training time, or FLOPs) or soften these claims to be qualitative/hypothesized rather than demonstrated properties.","section":"Section 5 and Abstract"},{"comment":"In Table 1, the set of vertices is written as V_KNN = {v1, d2, ..., dn}; the second element should presumably be v2. Please correct this typo.","section":"Table 1"},{"comment":"There are several typographical inconsistencies: 'Gumble-Softmax' should be 'Gumbel-Softmax', 'hope network' should be 'hop network', and Equation (5) uses 'P' in the denominator where 'Σ' would be clearer. These should be cleaned up before publication.","section":"Algorithm 1 and Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's novelty relative to DHGAT [46] is thin: the decision/representation split is essentially identical, with the action space changed from neighborhood type to hop order. This is not by itself disqualifying, but the omission of DHGAT and LOSS-GAT from the experimental comparison is a serious missing baseline issue, especially since those are the authors' own prior methods. The baseline adaptation ambiguity and unreported K choices are load-bearing and need to be resolved with a clean re-run. If the authors can supply a faithful comparison and report the relevant hyperparameters, the paper could be suitable for publication; as it stands, the central empirical claim is not yet verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the mechanism is plausible and the GATv2 ablation is the strongest part of the paper; the headline outperformance claim is not yet backed by a fair protocol. I would send it to peer review, but with major revision expected.\n\nThe method is a legitimate recombination: a hop network selects a neighbor distance per node per layer via Gumbel-Softmax, and an embedding network updates using that chosen hop. That is close to the authors' own DHGAT (decision network + representation network) plus L2P-style per-node propagation-depth learning, but applied to a homogeneous KNN graph with hop-order choices. The paper cites those predecessors, so the lineage is acknowledged rather than hidden. It is incremental work, not a breakthrough, and the abstract oversells it.\n\nThe GATv2 comparison in Figure 3, where both models share the same graph and architecture and differ only in message passing, gives real support to the core mechanism. That is the cleanest experiment in the paper. The problem is everything around it. TextGCN and LSTM-CP-GCN are described as following their original architectures, but they are actually reduced to a generic two-layer GCN on the same news-only KNN graph. TextGCN's mechanism is the word-document heterogeneous graph; LSTM-CP-GCN's mechanism is the sentence graph with CP-decomposed co-occurrence weights. Stripping those out makes the baselines weaker than their published designs, so the +0.19 margin over TextGCN on Fake.Br at 10% labels is suspect. The omission of DHGAT and LOSS-GAT, the closest methods from the same group, is a serious gap. The per-dataset K used for the main tables is never reported, even though Section 6.4.3 tunes K and finds an optimum at 6–7. If NOL-GAT's numbers were selected at the best K while baselines used a fixed default, the margins are inflated by test-set tuning. The over-squashing and computational-complexity claims are asserted without any measurement; they should be either demonstrated or removed.\n\nWho is this for? Researchers working on semi-supervised fake news detection and adaptive message-passing GNNs. The architecture is testable, and the code link is a point in the paper's favor. But I would not cite the numbers until the baselines are fixed and the K reporting is clarified.\n\nRecommendation: accept for peer review, but only with a strong referee who will demand a clean experimental protocol, comparisons to DHGAT/LOSS-GAT/L2P, and retraction or measurement of the complexity and over-squashing claims.","headline":"A coherent but incremental GNN architecture whose large reported gains over baselines rest on an unfair-looking baseline setup; worth refereeing, but only after a clean re-run.","tokens_in":21335,"tokens_out":2326,"would_cite":false,"duration_ms":22728,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph attention network that lets each news node choose its own k-hop neighborhood in every layer outperforms fixed-depth message passing for fake news detection, especially when labeled data is scarce.","keywords":["fake news detection","graph neural networks","semi-supervised learning","neighborhood order learning","Gumbel-Softmax","graph attention network","adaptive message passing","KNN similarity graph"],"falsifier":"Re-run the six baselines with their own published optimal hyperparameters and original feature pipelines on the same five datasets, add DHGAT and LOSS-GAT to the comparison, and check the 10%-label accuracy gap; if NOL-GAT's margin over the best proper baseline falls below roughly five accuracy points, or if either omitted method matches or beats it, the central outperformance claim is not supported.","tokens_in":20210,"feed_emoji":"📰","tokens_out":7349,"duration_ms":59490,"temperature":0.7,"pith_summary":"This paper tries to establish that a graph neural network can detect fake news better by letting every node—each news article—choose its own neighborhood distance in every layer, rather than being limited to the fixed L-hop reach of a standard L-layer network. The proposed NOL-GAT couples a Hop Network that picks a k-hop order for each node via Gumbel-Softmax with an Embedding Network that aggregates along that chosen order. Across five fake-news datasets and label fractions of 10%, 20%, and 30%, the paper reports accuracy and macro-F1 gains over six semi-supervised baselines, with the largest margins at 10% labels. A sympathetic reader would care because the mechanism addresses a known bottleneck—distant but semantically relevant neighbors are unreachable in shallow GNNs—without adding depth, and because low-label settings are where fake-news detection is most needed in practice.","feed_headline":"For fake news, letting each node pick its own neighborhood wins","feed_subtitle":"At 10% labeled data, it outperforms the best baseline by up to 19 accuracy points across five datasets.","key_machinery":"The load-bearing mechanism is the paired Hop Network Φ and Embedding Network Ψ, both implemented as GATv2 networks. Φ consumes each node's current embedding together with its φ-hop neighbors and outputs a probability vector over Γ; the Straight-Through Gumbel-Softmax turns that vector into a one-hot choice of neighborhood order γ, and Ψ then updates the node embedding from exactly the γ-hop neighbors. What this buys is a differentiable, node-level decision about reach: a node at the periphery of a KNN similarity graph can decide to listen to a 6-hop article, while a central node can stick to 1-hop or 2-hop neighbors, and the decision can change from layer to layer. The whole architecture operates on a fixed KNN graph built from Doc2Vec text embeddings, so the only difference from a standard GATv2 is this learned hop selection.","core_discovery":"The central claim is that fixed-depth message passing is the wrong constraint for semi-supervised fake news detection. NOL-GAT replaces the uniform rule 'aggregate over 1-hop neighbors in each layer' with a per-node, per-layer categorical choice: a probability distribution over neighborhood orders Γ = {0, 1, ..., dg}, where dg is the graph diameter, is output by a GATv2-based Hop Network, a neighborhood order is sampled differentiably with the Straight-Through Gumbel-Softmax estimator, and a second GATv2-based Embedding Network aggregates messages from exactly that hop distance. Because the selection is categorical and node-specific, the model can pull information from far-away nodes that a shallow standard GNN would never reach, while still avoiding the homogenization that deep stacking causes. The paper reports that this design outperforms all six compared baselines across five datasets and all three label proportions, and argues that it mitigates over-squashing and over-smoothing while keeping computational cost lower than deep alternatives.","pith_inferences":["If the per-node hop choices are stable across training runs, the sampled neighborhood orders could serve as a primitive explanation: which distant articles actually shaped a prediction. The paper does not analyze the learned distributions, so this is an extension, not a reported result.","The mechanism is graph-agnostic, so a natural next test is whether the same adaptive hop selection transfers to other semi-supervised node-classification tasks, such as citation or social-network label prediction.","A direct head-to-head with the closely related decision-based heterogeneous GAT (DHGAT) and the label-propagation method LOSS-GAT, both cited in the paper but absent from the experiments, would isolate whether the gain comes from hop-order flexibility or from other design choices such as heterogeneous neighbor types.","The complexity claim could be made quantitative by measuring wall-clock time and memory at matched accuracy against a deep GATv2; the paper asserts reduced complexity but does not report runtime."],"forward_implications":["At 10% labeled data, the reported margins over the best baseline reach about 19 accuracy points on Fake.Br and about 9 points on FakeNewsDetection, with similar macro-F1 gaps.","Because the hop choice is per layer, the same node can use a 4-hop neighborhood in layer 1 and a 2-hop neighborhood in layer 2, so the model adapts to local graph position rather than a global depth budget.","The reported optimal KNN graph parameter is k = 6 or 7 across datasets; smaller or larger values reduce macro-F1, showing that graph construction and hop selection interact.","NOL-GAT keeps the same architecture and hyperparameters as its GATv2 baseline except for the message-passing rule, so the paper attributes the gains specifically to adaptive neighborhood-order selection.","The model is content-only and operates on a similarity graph, so the claimed gains do not depend on user profiles or propagation trees, making the method applicable to any text dataset."],"supporting_citations":[{"why":"Supplies the Gumbel-Softmax estimator that makes discrete neighborhood-order selection differentiable.","marker":"[51]"},{"why":"GATv2 is the attention backbone for both the Hop Network and the Embedding Network.","marker":"[53]"},{"why":"Doc2Vec provides the 500-dimensional text embeddings from which the KNN graph and initial node features are built.","marker":"[47]"},{"why":"CO-GNN is the strongest dynamic-message-passing baseline and defines the action-selection paradigm NOL-GAT compares against.","marker":"[45]"},{"why":"DHGAT is the closest prior decision-network design from the same group; it is cited as related work but not included in the experiments.","marker":"[46]"},{"why":"TextGCN is the strongest classic GNN baseline on Fake.Br and provides the heterogeneous word-news graph baseline.","marker":"[58]"},{"why":"L2Q models per-node propagation steps and is the nearest adaptive-depth competitor tested.","marker":"[43]"}],"fun_headline_variants":["Each node picks its own neighborhood to spot fake news better","Adaptive neighborhoods boost fake news detection with few labels","Node-chosen hops beat fixed-depth GNNs for fake news","Learning ideal neighbor depth cuts fake news errors","When nodes choose their own scope, fake news detection jumps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline 'significantly outperforms' claim assumes the baselines were given a fair run—same Doc2Vec features and hyperparameters as NOL-GAT while faithfully preserving their original architectures—and that no omitted close competitor (DHGAT, LOSS-GAT) would match NOL-GAT's numbers.","fun_headline_variants_meta":{"raw":{"variants":["Each node picks its own neighborhood to spot fake news better","Adaptive neighborhoods boost fake news detection with few labels","Node-chosen hops beat fixed-depth GNNs for fake news","Learning ideal neighbor depth cuts fake news errors","When nodes choose their own scope, fake news detection jumps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1451,"prompt_tokens":985,"completion_tokens":466,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":387}},"tokens_in":601,"tokens_out":466,"duration_ms":4166,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:22:40.963621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the six baselines with their own published optimal hyperparameters and original feature pipelines on the same five datasets, add DHGAT and LOSS-GAT to the comparison, and check the 10%-label accuracy gap; if NOL-GAT's margin over the best proper baseline falls below roughly five accuracy points, or if either omitted method matches or beats it, the central outperformance claim is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gumbel-Softmax estimator that makes discrete neighborhood-order selection differentiable."},{"cited_title":"Brody, U","cited_arxiv_id":null,"evidence_quote":"GATv2 is the attention backbone for both the Hop Network and the Embedding Network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Doc2Vec provides the 500-dimensional text embeddings from which the KNN graph and initial node features are built."},{"cited_title":"Finkelshtein, X","cited_arxiv_id":null,"evidence_quote":"CO-GNN is the strongest dynamic-message-passing baseline and defines the action-selection paradigm NOL-GAT compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TextGCN is the strongest classic GNN baseline on Fake.Br and provides the heterogeneous word-news graph baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"L2Q models per-node propagation steps and is the nearest adaptive-depth competitor tested."}],"review_version":1}