{"id":"3fb402f7-15ca-4583-9104-644cd53d60d7","arxiv_id":"2508.01278","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FNGCN selects local centralities via a correlation-based feature network and uses them as GCN features, matching or slightly beating baselines on most but not all of six networks.","lead":"FNGCN is a graph neural network that groups eleven cheap local centrality measures into a 'feature network' by their correlations, picks one representative from each group, and feeds those into a GCN to decide which nodes are influential. Gains over simpler baselines are small, typically 0.1 to 1 percent, and the paper does not report error bars or statistical tests.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Feature-network selection is underspecified and its benefit is contradicted by the paper's own ablation; the claimed contribution is not established.","rationale":"I agree with the reader's weakest_assumption. The claim that FNGCN is a novel framework is load-bearing on the feature-network selection being both well-defined and beneficial. The paper gives no sensitivity analysis for δ, no specified community detection algorithm, and allows arbitrary tie-breaking; all three are explicit in Section III-B. The ablation is the decisive internal evidence: in Fig. 5, FNGCN sometimes underperforms GCN+LC, so the selection is not consistently helpful. Moreover, because no error bars or seeds are reported, the improvements in Table V might be within run-to-run variance. The concern is not that the method is fraudulent; it is that the paper's own evidence does not establish the mechanism claimed. A controlled sensitivity and significance re-run would settle it. If the selection is stable and beneficial, the paper's contribution stands; otherwise the method reduces to a standard GCN with local-centrality features, and the verdict should remain conditional pending code/data release.","tokens_in":20838,"tokens_out":5856,"duration_ms":68283,"concrete_test":"Run the full FNGCN pipeline on CA-CondMat with δ ∈ {0.6, 0.7, 0.8, 0.9, 0.95}, with two named community detection algorithms (e.g., Louvain and label propagation) for step S3, and with both tie-break choices in step S4; repeat FNGCN3 and FNGCN64 training with at least 10 random seeds and report mean ± std of accuracy, F1, and AUC. Also run the same GCN3+LC and GCN64+LC baselines under identical seeds. If the selected feature sets change with δ or community detection, or if FNGCN does not significantly outperform GCN+LC (e.g., paired t-test or Wilcoxon across seeds), then the feature-network step is not the source of the reported improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (Section III-B) is the only novel component separating FNGCN from a standard GCN with local-centrality features, but it is not fully specified: the Spearman threshold δ=0.9 is fixed with no sensitivity analysis, step S3 says 'any community detection algorithm' without naming one, and step S4 breaks ties arbitrarily ('we can choose either of them'). The selected features vary across networks (Table III), so the method is not deterministic as presented. More importantly, the paper's own ablation (Section IV-D, Fig. 5) shows FNGCN3 below GCN3+LC on CA-HepTh in accuracy and F1, and FNGCN64 below GCN64+LC on Human protein (Vidal) and CA-GrQc. Thus the feature-network selection does not consistently beat simply using all 11 local centralities as GCN features. Since 'use local centralities as GCN features' is already standard practice (e.g., InfGCN uses degree and clustering coefficient), the reported gains cannot be attributed to the feature network unless the selection step is shown to be stable and beneficial. No error bars, seeds, or sensitivity analysis are reported, so the small differences in Table V could be noise. The central claim that FNGCN 'can identify the influential nodes more accurately' therefore rests on an underspecified and empirically challenged selection mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FNGCN, a graph convolutional network framework for identifying influential nodes. It computes 11 local centrality features, builds a 'feature network' in which centralities are connected by Spearman correlation above a threshold, groups centralities via community detection, and selects one representative from each group as node features. Two GCN variants, shallow FNGCN3 and deep FNGCN64, are trained on SIR-model labels and evaluated on six real-world networks against SVM, LR, InfGCN, GATv2, and GATv2-FN. The paper also reports ablations on network depth, local versus global centralities, the benefit of centrality selection, and per-feature contributions. The central claim is that FNGCN identifies influential nodes more accurately than the compared methods while spending much less time on feature construction.","tokens_in":21036,"tokens_out":4873,"duration_ms":58012,"significance":"If the claims were established, the contribution would be moderate but useful: a GNN-based influence-identification method that uses only local centralities and a principled redundancy-selection step could reduce computational cost and improve interpretability. The manuscript has several good ingredients: the SIR ground truth is computed independently of the model, the ablation suite is broad, and the time-consumption comparison in Table VI supports the efficiency advantage of local centralities. However, the central comparative claim is not currently supported because the feature-selection algorithm is underspecified, the paper's own ablation does not consistently favor selection, and all numerical comparisons lack variance estimates. The work would be strengthened by a fully specified selection procedure, a sensitivity analysis, and controlled re-runs of all baselines under the same protocol.","major_comments":[{"comment":"Algorithm 1 is the only novel component separating FNGCN from a GCN with local-centrality features, but it is not fully specified. Step S3 says 'any community detection algorithm' without naming or parameterizing one; the Spearman threshold δ=0.9 is fixed with no sensitivity analysis; and step S4 breaks ties arbitrarily, as the text says 'we can choose either of them.' Because the selected feature sets differ across the six networks (Table III), the method is not deterministic or reproducible as presented. Please specify the community detection algorithm, state how δ was chosen, and report the stability of the selected features and downstream results with respect to δ and tie-breaking.","section":"Section III-B (Algorithm 1)"},{"comment":"The paper's own ablation contradicts the claim that feature-network selection is beneficial. In Fig. 5(a), FNGCN3 is below GCN3+LC on CA-HepTh in accuracy and F1; in Fig. 5(b), FNGCN64 is below GCN64+LC on Human protein (Vidal) and CA-GrQc. Since GCN+LC uses all 11 local centralities as features, the selection step does not consistently beat simply using no selection. This undermines the attribution of the Table V gains to the feature network. The authors should either provide a systematic comparison showing when selection helps or hurts and why, or qualify the contribution as comparable to using all local centralities with reduced feature dimensionality.","section":"Section IV-D, Fig. 5"},{"comment":"All reported accuracies, F1 scores, and AUC values are single point estimates, with no standard deviations, number of random seeds, or significance tests. Many differences are extremely small, on the order of 0.1 to 0.5 percentage points, so they could easily be noise. This is particularly problematic because the headline claim is that FNGCN identifies influential nodes 'more accurately' than the compared methods. Please rerun each configuration with multiple random seeds, report mean and standard deviation, and where possible include paired significance tests.","section":"Section IV-B/C, Tables V and Figs. 3-7"},{"comment":"The InfGCN results are transcribed from the original paper rather than re-run under the same training, labeling, and sampling pipeline, and the footnote to Table V concedes that accuracy values and all Sister-Cities results are not available from the original paper. Because FNGCN's SIR labels, 1:2 positive-negative sampling, and 70/30 train-test split may differ from the InfGCN setup, this baseline comparison is not controlled. Please re-run InfGCN under the protocol used for the other methods, or explicitly label the comparison as 'as reported in [6]' and avoid claiming superiority over it.","section":"Section IV-C, Table V"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'indentifying' in the Abstract, 'redundency', 'Condctance', 'nosies', 'aviod', 'baesd', and 'diffrent'; the manuscript needs a careful copyedit.","section":"Throughout"},{"comment":"Table I is heavily garbled by LaTeX and many formulas are unreadable, especially those for CoredCosine, CoredJaccard, and CoredPearson; the table should be regenerated so that every centrality definition is legible.","section":"Table I"},{"comment":"Equation (4), the Spearman correlation formula, is typeset incorrectly with misplaced parentheses and square roots, and the definitions of x_i and y_i should be stated more clearly; the current text uses 'ordinal number' but it is not clear whether ties are handled.","section":"Eq. (4)"},{"comment":"Reference [39] contains the placeholder 'http://www.address of you wannar cite/' and should be replaced with the proper KONECT dataset citations.","section":"Reference [39]"},{"comment":"The ablation numbering is inconsistent: Fig. 4's caption calls the local/global/all-centrality comparison the 'first ablation experiments,' while Section IV-D describes the hidden-layer study as the first ablation and later refers to Fig. 5 as the 'second ablation experiment'; the labels should be aligned.","section":"Section IV-D and figure captions"},{"comment":"The conclusion repeats the sentence 'identify the nodes with the best centrality values as influential nodes' twice; the duplicate phrase should be removed.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of cs.SI and the core idea is potentially publishable, but the evidence for the central claim is currently too weak: the selection algorithm is underspecified, the ablation data partly contradict the claimed benefit of selection, and no variance information is provided. I would encourage the editor to require a revised version with a fully reproducible selection procedure, sensitivity analysis, multi-seed results, and re-run baselines before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea—select local centralities as GCN features by clustering a Spearman correlation network—is coherent and arguably new, but the paper doesn't demonstrate that the selection step earns its keep. The experiments are extensive and the SIR labeling is done properly, independent of feature construction. The time-saving argument for local centralities is concrete and supported by Table VI.\n\nWhat the paper does well: the ablation structure actually asks the right questions (does depth matter, are local centralities enough, does selection help, which features matter). The SIR ground truth is standard and not circular with feature selection. The writing is clear enough to follow the pipeline.\n\nSoft spots, in rough order of severity. First, Algorithm 1 is underspecified: delta=0.9 is fixed with no sensitivity analysis, the community detection step just says 'any' without naming one, and step S4 breaks ties arbitrarily. As written, the selected features are not uniquely determined. Second, the paper's own ablation in Fig. 5 shows FNGCN3 below GCN3+LC on CA-HepTh accuracy/F1 and FNGCN64 below GCN64+LC on Human protein and CA-GrQc. That means the selection does not consistently beat using all eleven local centralities, so the headline claim that FNGCN identifies influential nodes 'more accurately' than compared methods is not supported by the selection mechanism. Third, there are no error bars, seeds, or multiple runs; many of the reported gains are fractions of a percent, so we can't tell if they're noise. Fourth, InfGCN numbers are transcribed from the original paper rather than re-run in the same pipeline, which makes the comparison uncontrolled. Fifth, no code or data release, and ref [39] has a placeholder URL.\n\nThe stress-test note is on target. I'd add that the abstract overclaims: 'more accurately than the compared methods' is too strong given SVM beats FNGCN on CA-HepTh accuracy and on CA-GrQc AUC.\n\nWho this is for: someone working on GNN-based influence identification might find the feature-network selection idea worth exploring, but the paper as it stands is not a solid empirical contribution. It deserves a serious referee because the idea is genuine and the experiments are extensive, but it needs major revision: specify the algorithm, run multiple seeds, re-run baselines under the same pipeline, and correct the overclaims. I'd engage with it as a borderline revise-and-resubmit, not a reject.","headline":"The feature-network selection idea is real but the paper's own ablation undermines the claim that it beats just using all local centralities.","tokens_in":21659,"tokens_out":1834,"would_cite":false,"duration_ms":20144,"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 GCN that selects node features by clustering correlated local centralities beats global-centrality baselines on influential-node detection.","keywords":["influential node identification","graph convolutional network","feature network","local centrality","Spearman correlation","SIR model","node classification","complex networks"],"falsifier":"A decisive test would run the selection with several thresholds $\\delta$ and several community-detection algorithms on the same six networks and same train/test splits; if accuracy and F1 vary strongly with those choices, or if a GCN fed all eleven local centralities matches or beats the selected subset, then the feature-network selection is not carrying the reported gains.","tokens_in":20511,"feed_emoji":"🕸️","tokens_out":7438,"duration_ms":81378,"temperature":0.7,"pith_summary":"The paper tries to show that the choice of node features is the limiting factor in graph-neural-network methods for finding influential nodes, and that a principled feature-selection step fixes it. FNGCN builds a small \"feature network\" over eleven local centrality measures, connects centralities whose Spearman correlation exceeds a threshold, clusters them, and picks one representative from each cluster to feed a GCN. Against SVM, logistic regression, InfGCN, GATv2, and GATv2-FN on six real-world networks, both the shallow and deep FNGCN variants achieve higher accuracy, F1, and AUC in most comparisons, and they spend far less time constructing features because global centralities are avoided. The paper also concludes that local centralities alone are sufficient, that the number of GCN layers matters little, and that Conductance of Egonet is the most informative feature.","feed_headline":"Clustered local features beat global centralities in GCN","feed_subtitle":"A GCN fed a small, de-correlated set of local centralities finds influential nodes more accurately and far faster.","key_machinery":"The feature network is the load-bearing object: each node is a local centrality, and an edge is placed between two centralities when their Spearman correlation coefficient is above $\\delta = 0.9$. Community detection on this small graph groups redundant centralities, and the selection rule keeps one centrality per group — the node with largest degree that has no already-chosen neighbor — giving a small feature set that is then rank-normalized into the node-feature matrix. The GCN receives this matrix together with the adjusted transition matrix $\\tilde{P} = \\tilde{D}^{-1/2}\\tilde{A}\\tilde{D}^{-1/2}$; the shallow variant stacks three layers with the standard convolution, while the deep variant uses the residual and identity-mapping update of [13] to stack 64 layers without over-smoothing.","core_discovery":"On the paper's own terms, the central claim is that representing the relationships among local centralities as a feature network and using that network to select a compact, de-correlated feature set improves a GCN's ability to identify influential nodes. The experiments show FNGCN3 and FNGCN64 outperforming the comparison methods on most of the six networks across accuracy, F1, and AUC, with the main degradations confined to CA-GrQc and CA-HepTh; ablation experiments are then used to argue that local centralities are sufficient, that deep layers give only marginal gains, that centrality selection is necessary, and that Conductance of Egonet contributes most to the model.","pith_inferences":["Editorial: because the selection step operates only on the input features and the network structure, the same framework could be dropped onto other GNN architectures or other node-classification problems; the paper only tests GCN.","Editorial: the fixed threshold $\\delta=0.9$ and the unspecified community-detection algorithm leave room for sensitivity analysis; sweeping the threshold would clarify how much of the gain comes from the particular selection rule.","Editorial: the prominence of Conductance of Egonet raises the possibility that a much simpler classifier using only that centrality plus degree would already achieve most of the accuracy, a baseline the paper does not report.","Editorial: because node features are rank-normalized, the method should transfer across networks of different sizes without rescaling; testing on directed or weighted networks would reveal whether the local-centrality definitions remain sufficient."],"forward_implications":["Feature construction drops from tens or hundreds of seconds to under one second on the six networks, because only local centralities are computed, so the method is practical on larger graphs.","A 3-layer GCN with selected local features is within about one point of a 64-layer GCN on most datasets, meaning expensive deep stacks are not required for this task.","Because the selected feature set changes from network to network (for instance, Degree is added for Human protein (Vidal) but not the others), the feature-network step is doing genuine per-network selection rather than applying a fixed list.","Conductance of Egonet is the single most important feature in the ablation, so future centrality-based classifiers for these networks should include it.","Local centralities alone match or beat global centralities in the comparisons, so the heavy cost of betweenness and closeness is avoidable."],"supporting_citations":[{"why":"Provides the InfGCN baseline whose global-plus-local feature scheme FNGCN is compared against and whose input features are reused in the ablation.","marker":"[6]"},{"why":"Supplies the graph convolutional layer that FNGCN3 uses for message passing.","marker":"[14]"},{"why":"Supplies the residual and identity-mapping update that allows the 64-layer FNGCN to avoid over-smoothing.","marker":"[13]"},{"why":"Supplies the SIR spreading model used to score every node and label the top 5 percent as influential.","marker":"[15]"},{"why":"Supplies the Spearman correlation coefficient used to determine edges in the feature network.","marker":"[35]"},{"why":"Supplies the GATv2 model used as a baseline, in its plain and feature-network-variant forms.","marker":"[44]"}],"fun_headline_variants":["Feature network trims centralities, GCN pinpoints influencers","De-correlated local centralities boost GCN accuracy","Local centrality feature graph speeds influencer detection","GCN with de-correlated centralities finds influencers faster","Feature graph selects local metrics for GCN, beats global ones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Algorithm 1's feature selection — the fixed Spearman threshold, the unspecified community-detection step, and the arbitrary tie-breaking — actually picks out the best local centralities, so that the reported accuracy gains come from the feature network rather than from the GCN itself or from per-network tuning.","fun_headline_variants_meta":{"raw":{"variants":["Feature network trims centralities, GCN pinpoints influencers","De-correlated local centralities boost GCN accuracy","Local centrality feature graph speeds influencer detection","GCN with de-correlated centralities finds influencers faster","Feature graph selects local metrics for GCN, beats global ones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3470,"prompt_tokens":968,"completion_tokens":2502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2424}},"tokens_in":584,"tokens_out":2502,"duration_ms":20167,"temperature":1.0,"reasoning_tokens":2424,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:43:45.465850+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would run the selection with several thresholds $\\delta$ and several community-detection algorithms on the same six networks and same train/test splits; if accuracy and F1 vary strongly with those choices, or if a GCN fed all eleven local centralities matches or beats the selected subset, then the feature-network selection is not carrying the reported gains.","supporting_citations":[{"cited_title":"Infgcn: Identifying influential nodes in complex networks with graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Provides the InfGCN baseline whose global-plus-local feature scheme FNGCN is compared against and whose input features are reused in the ablation."},{"cited_title":"Simple and deep graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the residual and identity-mapping update that allows the 64-layer FNGCN to avoid over-smoothing."},{"cited_title":"Mathematical modeling of diseases: Susceptible-infected-recovered (sir) model,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIR spreading model used to score every node and label the top 5 percent as influential."},{"cited_title":"Spearman correlation coefficients, differ- ences between,","cited_arxiv_id":null,"evidence_quote":"Supplies the Spearman correlation coefficient used to determine edges in the feature network."},{"cited_title":"How attentive are graph attention networks?","cited_arxiv_id":null,"evidence_quote":"Supplies the GATv2 model used as a baseline, in its plain and feature-network-variant forms."}],"review_version":1}