{"id":"7c60d6e8-dd55-4b65-b6bd-ee41902f2170","arxiv_id":"2412.01949","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A network science paper introduces Smart Bins, a clustering-based label generation method, and two new prediction targets (influence peak and peak time), reporting improved and more stable key node classification.","lead":"The paper proposes an improved machine learning pipeline for identifying influential nodes in networks, using clustering to create training labels called Smart Bins. It also shows models can predict how fast and how high a spread peaks, albeit with open questions about how the advantage is measured.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Smart Bins advantage is measured by macro-F1 on labels that Smart Bins itself constructs; until an external key-node benchmark is run, the central 'higher and more stable results' claim is not established.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the evaluation of Smart Bins is circular in the sense that macro-F1 is computed on labels generated by the very discretization schemes being compared, and these schemes differ in class count, class balance, and threshold placement. My reading of the manuscript confirms this. Section 3.2 states that KMeans is run on all node influence ranges and that the number of bins is selected based on membership counts, while the competing method is a binary top-5% split. Section 4.3 then compares these under macro-F1 without any external validation of whether the predicted top class corresponds to nodes that actually generate larger cascades. This is the central contribution's main empirical support, so the concern is genuinely load-bearing. I do not see an internal inconsistency or a fraudulent claim; the framework itself is coherent, and the near-perfect same-network results in Section 4.1 are plausible given the strong correlation between the selected centrality features and IC influence. But the comparison in Figure 6 cannot establish the claimed improvement without a benchmark that is independent of the labeling procedure. The proposed concrete test—using seed-set IC spread or NDCG against true influence—would settle whether the macro-F1 advantage reflects a real identification gain. Because the issue is addressable and the core pipeline may survive such a test, the appropriate verdict remains CONDITIONAL, matching the reader's assessment.","tokens_in":11496,"tokens_out":3777,"duration_ms":35581,"concrete_test":"Re-run the Section 4.3/Figure 6 comparison with an external, label-independent metric. For each labeling scheme and network, train LightGBM on the same train/test split, then rank all held-out nodes by the predicted probability of belonging to the highest (most influential) class. Select the top k nodes, where k is the number of nodes in the ground-truth top 5% by true IC influence, and compare (a) the expected IC spread of the selected seed sets and (b) NDCG@k of the ranking against true influence values, between Smart Bins and the fixed top-5% baseline. If Smart Bins' advantage persists on these external metrics, the macro-F1 gain is not merely a label-construction artifact; if it shrinks or reverses, the paper's central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, repeated in Section 4.3 and the Discussion, is that Smart Bins 'significantly improves the inference process in terms of higher and more stable results' compared with fixed binning such as the arbitrary top-5% approach. The only evidence for this is macro-F1 on node classification, but the comparison is confounded. Smart Bins chooses the number of bins (between 2 and 5, Section 3.2) and fits KMeans on the same ground-truth influence values it later uses to label; the fixed baseline is a binary top-5% split. Thus the two schemes differ simultaneously in label cardinality, class balance, and threshold positions. Macro-F1 is sensitive to all three: a 5-class, relatively balanced labeling problem is not directly comparable to a highly imbalanced binary problem, and a higher F1 may merely reflect that Smart Bins labels are easier to predict, not that the identified key nodes are better. No external benchmark—such as the actual IC spread of selected seed sets, top-k ranking quality, or NDCG against ground-truth influence—is used. This assumption is load-bearing because Section 4.3/Figure 6 is the paper's main empirical case for Smart Bins. The cross-network generalization results in Section 4.2 add a second concern: class labels are network-specific, since KMeans thresholds are fit per network, so the same label can denote different influence ranges across networks. The primary issue remains that the headline comparison is computed on self-constructed labels, and the claimed advantage may be an artifact of the evaluation protocol rather than a real improvement in key node identification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine learning framework for identifying key nodes in complex networks under the Independent Cascade model. The main contribution is 'Smart Bins', a KMeans-based discretization of simulated influence range values that produces class labels for supervised node classification. The paper also introduces two new prediction targets—influence peak and peak time—and evaluates a LightGBM classifier on four real-world networks (Citeseer, Pubmed, Facebook, Github). The authors report near-perfect performance in a within-network train/test split, moderate cross-network generalization, and claim that Smart Bins significantly outperforms the fixed top-5% binning approach used in prior work.","tokens_in":11780,"tokens_out":4857,"duration_ms":49432,"significance":"If fully supported, the framework would offer a fast and accurate way to estimate node influence classes and additional spreading characteristics without rerunning expensive IC simulations. The paper has concrete strengths: it provides a reproducible pipeline (code on GitHub), uses multiple real-world networks of different types and sizes, and includes a broad feature importance analysis with Shapley values. The proposed new tasks (influence peak, peak time) are a useful extension of the key-node identification problem. However, the central comparison that Smart Bins 'significantly improves the inference process' is currently not convincing: the evaluation relies on macro-F1 computed on labels that Smart Bins itself constructs, and the comparison to the fixed top-5% baseline is confounded. The cross-network generalization results are also difficult to interpret because class labels are not aligned across networks.","major_comments":[{"comment":"The comparison between Smart Bins and the fixed top-5% binning is not a valid head-to-head because the two schemes differ simultaneously in label cardinality (Smart Bins uses 2–5 classes, the baseline is binary), class balance, and threshold positions. Macro-F1 is sensitive to all three; a higher value for Smart Bins may simply reflect that the KMeans-induced labels are easier to predict, not that the identified key nodes are better. The paper's central claim of 'higher and more stable results' is therefore not established. Please add an external evaluation metric that directly measures key-node identification quality, such as the average IC spread of the top-k predicted seeds, top-k precision or recall against the ground-truth top spreaders, or NDCG against the true influence range. The comparison should also be made under matched conditions—for example, the same number of classes or the same class balance.","section":"Section 4.3, Figure 6"},{"comment":"KMeans is fit on the influence values of all nodes, including the held-out test nodes, before the train/test split. Although the classifier does not observe test labels in training, the label definition itself is derived from the full data distribution. This makes the classification task easier than a realistic setting where label thresholds must be inferred from the training set only (or from a validation set). Please either derive label boundaries using only training data and then apply them to test nodes, or justify the current procedure and discuss its effect on the reported near-perfect F1 scores.","section":"Section 3.2"},{"comment":"The cross-network generalization results are reported with class labels that are not aligned across networks, because KMeans thresholds are fit per network. The same label index (e.g., 3) represents different influence ranges in different networks. This makes the reported F1 scores and the conclusion that 'the family of the network matters more than its size' difficult to interpret, since the classifier is effectively predicting different target definitions on the training and test networks. Please align labels across networks—for instance, by fitting thresholds on the training network and applying them to the test network, or by using quantile-based thresholds that are defined consistently—and re-run the generalization experiments.","section":"Section 4.2, Figure 5"},{"comment":"The claim that Smart Bins provide 'significantly more stable results' is not supported by any statistical test or confidence interval. Figure 6 appears to show box plots, but the text does not report the number of repeated trials, the variance, or any significance test. Please report the distribution of the results across runs and test whether the difference in stability is statistically significant (e.g., with a paired test across the same node splits).","section":"Section 4.3"}],"minor_comments":[{"comment":"The KMeans formula as written, sum over i of min over mu_j of ||x_i - mu_j||_2, is not the standard KMeans objective; the within-cluster sum of squares should be a double sum over clusters and their members, and the squared norm is usually used. Please rewrite the objective correctly.","section":"Section 3.2"},{"comment":"The sentence 'we assigned each cluster member a centroid (cluster centre) value' is unclear: presumably the authors assign a cluster label, not the centroid value itself. Please rephrase.","section":"Section 3.2"},{"comment":"The table reports the percentage of nodes whose influence range values fall into the 'top bin' for 2–5 KMeans bins, but it is not stated how the top bin is defined—by the highest centroid, or by the bin containing the maximum influence range? Please clarify.","section":"Table 1"},{"comment":"The activation probability thresholds are given as sets (0.2, 0.3, 0.4 for citation networks; 0.1, 0.15, 0.2 for social networks), but the paper does not state whether the reported results are averaged over these thresholds or reported per threshold. Please clarify the aggregation.","section":"Section 3.1"},{"comment":"The y-axis label is missing; it is presumably the macro-F1 score. Please add axis labels to all figures that lack them.","section":"Figure 3"},{"comment":"The statement that the SIR model with recovery rate 1.0 'effectively reduces to the IC model' is correct, but a citation or a brief derivation would help the reader, especially since this equivalence is used to justify the choice of the IC model.","section":"Section 5, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of Entropy and addresses a timely problem, but the current evaluation of Smart Bins is circular: the headline metric is macro-F1 on labels that the method itself generates, and the baseline comparison is confounded. I believe the central claim can be rescued with additional experiments that use an external key-node benchmark and aligned labels across networks. If the authors cannot provide such validation, the main contribution would be reduced to a description of a clustering-based discretization procedure rather than a demonstrated improvement. I would also encourage the editor to consider whether the near-perfect within-network F1 scores (Figure 4) are meaningful without external validation of the label quality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper with two genuinely new things — Smart Bins and the peak/peak-time prediction targets — but the main claim that Smart Bins beats fixed binning is not actually shown. The only evidence is macro-F1 on labels Smart Bins itself constructs, so the comparison is confounded.\n\nWhat's good: The framework is clearly described, the IC simulations are done properly (100 runs per node per threshold), and the authors test on four real networks. The feature importance analysis (SHAP) is a nice addition, and the finding that out-degree, average neighbour degree and local reaching dominate is sensible. The two new prediction targets — peak size and peak time — are practically useful and, as far as I know, genuinely new to the key-node literature. The authors also promise code.\n\nThe soft spot is the evaluation of Smart Bins in Section 4.3/Figure 6. They compare Smart Bins to a binary top-5% fixed bin, and measure macro-F1 of a downstream classifier. Smart Bins chooses K and fits KMeans on the same ground-truth influence values it later labels; the fixed baseline is one arbitrary threshold. So three things differ at once: number of classes, class balance, and where the boundaries sit. A higher F1 on Smart Bins labels could just mean those labels are easier to predict, not that the method identifies better key nodes. On top of that, KMeans is fit on all nodes including the test set, so the test labels are not independent. The cross-network results in Section 4.2 add a second issue: class thresholds are network-specific, so the same label means different influence ranges across networks. None of this invalidates the framework, but it does mean the paper's central claim — 'significantly improves the inference process' — is not established.\n\nThese are fixable. I'd want an external benchmark: e.g., compare the top-k nodes selected by the classifiers against actual IC spread, or compute NDCG between predicted classes and ground-truth influence. And fit KMeans only on the training set. If the authors do that, the Smart Bins claim might hold up — the idea has merit because it adapts bin boundaries to the data. But without an external benchmark, the current evidence is circular.\n\nMy recommendation: send it to review, because the new targets and framework are worth discussing, and a good referee could push the authors to run the right experiments. But it should not be accepted as-is.\n\nWho's it for: network scientists working on ML-based influence prediction, and practitioners who need fast approximations of cascade dynamics. It's a useful paper to know about, but not one I'd cite for the Smart Bins advantage yet.","headline":"Competent engineering with two new ideas, but the Smart Bins claim rests on a confounded evaluation; needs an external benchmark.","tokens_in":12348,"tokens_out":3088,"would_cite":false,"duration_ms":226014,"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":"The paper claims that clustering simulated influence values into Smart Bins gives machine learning models a more accurate and stable way to identify key spreaders than fixed top-5% binning.","keywords":["social networks","node classification","unsupervised learning","influence spread","Independent Cascade model","key node identification","machine learning","centrality measures"],"falsifier":"Run both labeling schemes on the same networks, then from each scheme's top class select a fixed-size seed set and measure the actual Independent Cascade spread those seeds produce; if the top-5% seeds match or exceed Smart Bins seeds in realized spread, the claimed advantage is contradicted.","tokens_in":11230,"feed_emoji":"🕸️","tokens_out":3914,"duration_ms":37769,"temperature":0.7,"pith_summary":"This paper is trying to establish that how you turn simulated influence values into training labels matters as much as the classifier itself. It introduces Smart Bins, a KMeans-based discretization of nodes' mean influence under the Independent Cascade model, and argues this beats the arbitrary top-5% split used in earlier work. The payoff is that a machine learning model trained on cheap centrality features can reproduce the expensive simulation labels nearly perfectly on unseen nodes of the same network. The paper also claims the framework can predict not just how far influence spreads, but when it peaks and how large the peak is. If true, this would make key-node identification faster, cheaper, and more operationally useful for tasks like viral marketing and epidemic planning.","feed_headline":"Smart Bins beat the top 5 percent cutoff for influencer labels","feed_subtitle":"KMeans-based labeling makes ML models predict influence spread, peak, and peak time more reliably.","key_machinery":"Smart Bins is the central object: KMeans clustering applied to the one-dimensional array of nodes' mean simulated influence values, where each resulting cluster becomes a labeled class and the number of clusters is adjusted so every class has enough members. This is paired with a feature embedding of fourteen centrality measures—out-degree, average neighbour degree, local reaching, betweenness, PageRank, and others—plus the activation probability used in the diffusion model, all standardized and fed into a LightGBM classifier. The KMeans step carries the paper's main novelty, while the centrality features keep the inference cheap compared with rerunning simulations.","core_discovery":"The paper's central claim is that the process of obtaining training labels—not just the choice of machine learning algorithm—determines how well influential nodes can be identified. It introduces Smart Bins, which applies KMeans clustering to the distribution of nodes' average simulated influence values, with the number of bins chosen so each class has enough members, and replaces the arbitrary top-5% threshold used in prior work. Using these labels, a LightGBM classifier trained on fourteen centrality features plus the activation threshold reproduces the influence-range classes of expensive simulations almost perfectly on held-out nodes, and it can also predict the size of the peak cascade and the time required to reach that peak. Cross-network experiments show that models generalize best when training and test networks belong to the same family—citation to citation or social to social—suggesting that network type matters more than network size.","pith_inferences":["Because Smart Bins fits the labels to the same simulation values it labels, part of the reported F1 gain may reflect label construction rather than better key-node identification; a direct comparison of seed-set influence under a fixed budget would test this.","The paper evaluates classification quality with macro-F1, but a ranking-based evaluation (e.g., how well the top predicted class matches the top spreaders by quantile) might change the conclusions about generalization.","The result that network family matters more than size suggests that a network-similarity measure could be used to select small training networks for large targets, but the paper does not test such a measure.","The framework's reliance on KMeans implies that networks with heavily skewed influence distributions may need very different bin counts, so an adaptive rule for choosing the number of bins would be a natural extension."],"forward_implications":["Machine learning classifiers can reproduce the Smart Bins labels of expensive Independent Cascade simulations almost perfectly on held-out nodes of the same network.","The framework predicts not only total influence range but also peak cascade size and time to peak, giving operational forecasts for viral marketing and epidemic response.","Cross-network generalization works best within the same network family, so a smaller but topologically similar training network can outperform a larger but different one.","Smart Bins produces more stable results across repeated trials than an arbitrary top-5% fixed binning, because the bins adapt to the actual distribution of influence values.","Out-degree, average neighbour degree, and local reaching are the most informative centrality features for predicting influence spread."],"supporting_citations":[{"why":"Defines the Independent Cascade model used to compute ground-truth influence values.","marker":"[5]"},{"why":"Provides the arbitrary top-5% labeling baseline and the prior machine-learning approach Smart Bins is compared against.","marker":"[8]"},{"why":"Proposes a machine-learning framework with fixed uniform/quantile discretization, the labeling approach Smart Bins extends and improves.","marker":"[10]"},{"why":"Supplies the KMeans algorithm that Smart Bins uses for discretization.","marker":"[25]"},{"why":"Supplies the Citeseer and Pubmed citation-network datasets used in the experiments.","marker":"[38]"},{"why":"Supplies the Facebook and Github social-network datasets used in the experiments.","marker":"[39]"}],"fun_headline_variants":["Smart Bins: better influencer labels for ML spread models","KMeans-based labeling improves ML influence prediction","ML with Smart Bins predicts influence, peak, and peak time","Influence spread ML generalizes by network type, not size"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's main comparison assumes that macro-F1 scores from different labelings (Smart Bins vs. fixed top-5%) measure the same quality, even though the labels differ in number, balance, and threshold positions.","fun_headline_variants_meta":{"raw":{"variants":["Smart Bins: better influencer labels for ML spread models","KMeans-based labeling improves ML influence prediction","ML with Smart Bins predicts influence, peak, and peak time","Influence spread ML generalizes by network type, not size"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000524,"raw_usage":{"total_tokens":2524,"prompt_tokens":926,"completion_tokens":1598,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1530}},"tokens_in":542,"tokens_out":1598,"duration_ms":12250,"temperature":1.0,"reasoning_tokens":1530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:59:56.614953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both labeling schemes on the same networks, then from each scheme's top class select a fixed-size seed set and measure the actual Independent Cascade spread those seeds produce; if the top-5% seeds match or exceed Smart Bins seeds in realized spread, the claimed advantage is contradicted.","supporting_citations":[{"cited_title":"Maximizing the spread of influence through a social network","cited_arxiv_id":null,"evidence_quote":"Defines the Independent Cascade model used to compute ground-truth influence values."},{"cited_title":"Top influencers can be identified universally by combining classical centralities","cited_arxiv_id":null,"evidence_quote":"Provides the arbitrary top-5% labeling baseline and the prior machine-learning approach Smart Bins is compared against."},{"cited_title":"A machine learning based framework for identifying influential nodes in complex networks","cited_arxiv_id":null,"evidence_quote":"Proposes a machine-learning framework with fixed uniform/quantile discretization, the labeling approach Smart Bins extends and improves."},{"cited_title":"Least squares quantization in PCM","cited_arxiv_id":null,"evidence_quote":"Supplies the KMeans algorithm that Smart Bins uses for discretization."},{"cited_title":"Revisiting semi-supervised learning with graph embeddings","cited_arxiv_id":null,"evidence_quote":"Supplies the Citeseer and Pubmed citation-network datasets used in the experiments."},{"cited_title":"Multi-scale attributed node embedding","cited_arxiv_id":null,"evidence_quote":"Supplies the Facebook and Github social-network datasets used in the experiments."}],"review_version":1}