{"id":"7e0a9881-1706-462d-8def-e3346c8c6df9","arxiv_id":"2411.17138","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HGC combines effective-distance gravity, structural-hole constraint, and cycle-ratio propagation into one centrality score; on eight networks it is best on five, tied on some, and beaten by LGM on three.","lead":"This paper proposes HGC, a hybrid ranking score that blends a gravity model with effective distance, structural holes, and cycle structure to identify influential spreaders in networks. Tests on eight real-world networks show it is competitive with, but not uniformly better than, the local gravity model it extends.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract overstates HGC: Table 2 shows LGM outperforms HGC on USAir, Yeast, and Power, and no significance testing supports the remaining 0.005–0.018 margins.","rationale":"Focusing on the unspecified cycle-number-matrix algorithm is important for reproducibility, but it does not settle whether the advertised conclusion holds: even with exact cycle counts, the claim could still fail on the presented evidence. The paper's own Table 2 gives an internal reason to doubt the abstract: LGM beats HGC on three of eight networks, and the margins elsewhere are small. Because the reference rankings come from stochastic SIR simulations and no uncertainty quantification is reported, the observed differences cannot currently be distinguished from noise. A single re-analysis of Table 2 with confidence intervals and beta sensitivity is therefore the most direct test of the central claim. The method is coherent and the hybrid idea is plausible, so outright rejection is not warranted; the reader's conditional verdict remains appropriate, provided the headline claim is narrowed or the statistical evidence is strengthened.","tokens_in":12190,"tokens_out":5625,"duration_ms":55052,"concrete_test":"Compute bootstrap 95% confidence intervals for the Table 2 Kendall-tau values by resampling the 1000 SIR runs per network, or re-running with 10,000 runs, at beta = beta_c as well as at 0.5 beta_c and 1.5 beta_c. Apply a per-network permutation test comparing HGC against LGM. If the five favorable margins are within noise, or LGM remains significantly higher on USAir/Yeast/Power, the abstract's universal outperformance claim should be withdrawn in favor of a conditional or competitive claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive weakness is the evidentiary basis of the headline claim, not only the under-specified cycle-count computation. Section 5.1/Table 2 reports Kendall's tau at a single epidemic parameter beta = beta_c. HGC is not the best method on 3 of 8 networks: LGM is higher on USAir (0.7758 vs 0.766), Yeast (0.8099 vs 0.8085), and Power (0.7383 vs 0.7145). On the five networks where HGC leads, the margin over LGM is only 0.005–0.018, and the SIR-derived reference rankings are stochastic averages; the paper gives no confidence intervals, standard deviations, or significance tests. Section 5.1 acknowledges the Power outcome, but the abstract and conclusion assert unqualified outperformance. With R and T fixed to 2, no ablation, and no beta-sensitivity analysis, the claimed superiority is not established. The central claim would require either a narrowed statement (competitive with, or best on 5 of 8) or statistically tested margins.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HGC, a hybrid centrality for identifying influential spreaders in undirected unweighted networks. HGC combines a gravity-model interaction term using effective distance and structural-hole constraint with an information-propagation term based on cycle ratio, fused through a balancing factor. The method is evaluated on eight real-world networks using SIR simulations and compared with seven baselines in terms of Kendall's tau, Jaccard similarity, and monotonicity. The authors claim that HGC outperforms all seven compared methods.","tokens_in":12421,"tokens_out":6394,"duration_ms":61170,"significance":"If the claims were fully supported, HGC would be a useful contribution to influential-spreader identification, since it combines asymmetric effective-distance information with cycle-structure information rather than fitting free parameters to SIR outcomes. The paper has strengths: evaluation on multiple real networks, comparison with several established baselines, and no circular fitting of the target result. However, the headline claim of universal superiority is contradicted by the paper's own table, and the method's core interaction term appears to be disabled by the chosen truncation radius as written. These issues must be resolved before the contribution can be assessed.","major_comments":[{"comment":"The opening claim that HGC outperforms all seven compared methods is not supported by Table 2. LGM attains higher Kendall's tau on USAir (0.7758 vs 0.766), Yeast (0.8099 vs 0.8085), and Power (0.7383 vs 0.7145), and on the five networks where HGC leads, the advantages over LGM are between 0.005 and 0.018. The SIR reference values are stochastic averages of 1000 runs, yet no confidence intervals, standard deviations, or significance tests are reported. The abstract and conclusion should either be narrowed to \"competitive with\" or \"best on five of eight,\" or the margins need statistical support.","section":"Abstract; Section 5.1, Table 2"},{"comment":"The truncation radius R=2 appears incompatible with the effective-distance definition. From Eq. (1), the effective distance from a source i of degree k_i to any adjacent node j is 1 - log2(1/k_i) = 1 + log2(k_i), which exceeds 2 whenever k_i > 2. Therefore, under the condition ED(j|i) ≤ R in Eq. (5), GM(i)=0 for all nodes of degree greater than 2, and the gravity/interaction component contributes nothing for typical nodes in the datasets (e.g., USAir has average degree 12.8). Please clarify whether R is a hop count or an effective-distance threshold; if the latter, the experiments need to be redone with a meaningful R, and if the former, Eq. (5) must be revised.","section":"Section 3.1.1, Eq. (5)"},{"comment":"The computation of the cycle number matrix c_ij is not specified. The paper only states that c_ij counts shortest cycles through i and j, but gives no algorithm, data structure, or complexity bound. On Facebook (4039 nodes) and Power (4941 nodes), enumerating all shortest cycles through every pair is nontrivial; if an approximation or a cycle basis is used instead, the reported rankings depend on an unspecified choice. A complete and reproducible specification of this step is required.","section":"Section 3.1.2, Eq. (8), Algorithm 1"},{"comment":"No sensitivity or ablation analysis is provided for the fixed choices R=2, T=2, and the balancing factor gamma. Since gamma is set as a ratio of averages and is not tuned to the SIR outcome, this is not circular, but the narrow margins in Table 2 leave open the possibility that the results depend heavily on these choices. Reporting results for R=1,2,3 and T=1,2,3, together with an ablation that removes the GM or RCP term, would substantiate the claimed contribution of each component.","section":"Section 5, Tables 2 and 3"}],"minor_comments":[{"comment":"The Jaccard similarity coefficient is defined for sets, but the text refers to sequences; please clarify that the top-k lists are converted to sets and state the range of k used in Figure 4.","section":"Section 4.4, Eq. (11)"},{"comment":"The text says RDP achieved the best performance in the Email network and that RDP and LGM attained the highest values in PB and Facebook, but Table 3 shows HGC tied for first in all of those networks; please correct the description.","section":"Section 5.3, Table 3"},{"comment":"The axis labels and legends in Figures 4 and 5 appear garbled in the manuscript version; please provide high-resolution figures with readable text.","section":"Figures 4 and 5"},{"comment":"Line 12 computes the balancing factor gamma inside the neighbor loop; this should be computed once after the GM and RCP accumulations are complete.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is reasonable: the gravity model treats node pairs symmetrically, and replacing topological distance with Brockmann–Helbing effective distance is a sensible way to break that symmetry. Adding cycle ratio and a structural-hole constraint is a plausible combination. The paper does a fair job of situating itself in the literature and the method is clearly defined. That is the good part.\n\nThe soft spot is the headline claim. The abstract says HGC outperforms all seven compared methods, but Table 2 shows LGM winning on USAir, Yeast, and Power. On the five networks where HGC leads, the margins over LGM are 0.005–0.018 in Kendall's tau. No error bars, standard deviations, or significance tests are reported. SIR is stochastic; the averages over 1000 runs can easily shift by more than 0.01, so the remaining margins are not established. The conclusion and abstract should be narrowed to something like \"competitive with, and often better than, seven baselines\" unless the authors add proper statistical testing.\n\nThe other real gap is reproducibility. The cycle number matrix c_ij is used in Eq. 8, but no algorithm or complexity bound is given. Enumerating all shortest cycles on Facebook (4039 nodes) and Power (4941 nodes) is nontrivial. If an approximation or cycle basis is used, the rankings depend on an unspecified choice. The paper also fixes R=2 and T=2 with no ablation, and no code or data are released. These are not fatal flaws, but they are exactly what a serious referee would ask for.\n\nI do not think the work is fundamentally broken. The method itself is coherent, the effective-distance idea has merit, and the monotonicity results are uniformly strong. The circularity burden is low; gamma is defined as a ratio of averages, not fit to SIR outcomes. The main problem is the mismatch between the claim and the evidence.\n\nThis paper is for readers in the influential-spreader subfield who want a new hybrid ranking to compare against. It deserves peer review because the method is plausible and the question is relevant, but it needs a major revision: report significance, narrow the claims, specify the cycle-count computation, and release code. I would not cite it in my own work until those issues are addressed.","headline":"A coherent hybrid centrality that is competitive with, not clearly better than, LGM; the abstract overstates what Table 2 shows.","tokens_in":12962,"tokens_out":932,"would_cite":false,"duration_ms":9833,"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 paper proposes HGC, a centrality score that combines a gravity model with effective distance and cycle structure, and reports that it identifies influential spreaders more accurately than seven comparison methods on eight real-world…","keywords":["complex networks","influential spreaders","gravity model","effective distance","cycle structure","SIR model","centrality ranking","hybrid method"],"falsifier":"Take a tree network (no cycles at all): HGC's RCP term is zero for every node, so HGC collapses to its gravity term; if HGC then still outperforms LGM by the same margin it shows on cyclic networks, the cycle component is not the source of the claimed advantage. Alternatively, rerun the Facebook comparison with a different shortest-cycle enumeration and check whether the top-k list changes materially.","tokens_in":11968,"feed_emoji":"🕸️","tokens_out":6990,"duration_ms":59671,"temperature":0.7,"pith_summary":"The paper sets out to establish that a node's capacity to spread influence is better captured by combining the gravity model's mass-and-distance logic with an asymmetric distance and with cycle-structure information. To do this it proposes the HGC score, which sums a gravity interaction term $GM(i)$ and a cycle-ratio propagation term $RCP(i)$, scaled by a balancing factor $\\gamma$. On eight real-world networks, with SIR simulations as the ground truth for spreading, HGC achieves higher Kendall's tau values than degree, betweenness, closeness, k-shell, cycle ratio, local gravity, and RDP, and it also improves top-k Jaccard similarity, the spreading power of its top nodes, and ranking monotonicity. The paper concludes that replacing topological shortest-path distance with effective distance fixes the symmetry flaw of prior gravity models, and that cycle structure adds redundant-pathway information that matters in densely connected networks.","feed_headline":"Hybrid gravity-cycle score tops seven methods for finding key spreaders","feed_subtitle":"HGC folds asymmetric distance and cycle redundancy into one ranking; SIR tests on eight real networks favor it.","key_machinery":"The machinery is a two-term centrality. First, a gravity term $GM(i)=\\sum_{ED(j|i)\\le R} e^{-c(i)} k_i k_j / ED_{j|i}^2$, where $c(i)$ is Burt's network constraint coefficient for structural holes, $k$ is degree, and effective distance $ED_{j|i}=1-\\log_2(a_{ij}/k_i)$ replaces shortest-path distance. Second, a propagation term $RCP(i)=\\sum_{l=1}^{T} I_l(i)$ with $I_l(i)=l^{-2}\\sum_{j\\in\\Lambda_i} I_{l-1}(j)$, seeded by the cycle ratio $CR(i)$, which quantifies involvement in other nodes' shortest cycles. The balancing factor $\\gamma=\\langle GM\\rangle/\\langle RCP\\rangle$ fuses the two scores. Taken together, the machinery encodes three ideas: influence decays with probability-based asymmetric distance, bridging positions (structural holes) add weight, and nodes embedded in cycles inherit propagation power from their neighbours.","core_discovery":"The central discovery is a single score, $HGC(i) = GM(i) + \\gamma RCP(i)$, where $GM(i)$ accumulates $e^{-c(i)} k_i k_j / ED_{j|i}^2$ over neighbours within a radius $R$, $RCP(i)$ spreads each node's cycle ratio through $l^2$-damped neighbour iterations, and $\\gamma = \\langle GM\\rangle/\\langle RCP\\rangle$ balances the two scales. The effective distance $ED_{j|i}=1-\\log_2 P_{j|i}$, with $P_{j|i}=a_{ij}/k_i$, makes the distance from $i$ to $j$ different from the distance from $j$ to $i$, so influence between a pair is no longer symmetric. The cycle ratio $CR(i)$ counts how strongly node $i$ participates in the shortest cycles of neighbouring nodes, and the RCP iteration carries that structural information to higher-order neighbours. The paper argues that these two ingredients, asymmetric effective distance and cycle-based propagation, are what let HGC identify nodes that SIR spreading actually starts from better than the seven baselines.","pith_inferences":["The paper gives no algorithm or complexity bound for computing the cycle-number matrix $c_{ij}$, so on large networks such as Facebook the reported rankings depend on an unspecified shortest-cycle enumeration; this leaves the method's practical scalability open.","The asymmetric effective distance could be carried directly into directed or weighted networks, where influence asymmetry is the norm, though the paper's experiments are limited to undirected unweighted graphs.","A natural ablation would compare HGC against its own gravity-only term: if most of the Kendall-tau gain survives without $RCP(i)$, then cycle structure is not the decisive ingredient the narrative suggests.","The balancing factor $\\gamma$ is data-dependent, and on near-tree networks where cycle ratios vanish the RCP term collapses, so behaviour on very sparse networks is an untested regime."],"forward_implications":["HGC gives more accurate full-network rankings than the seven baselines on five of eight test networks, with second-best performance on the sparse Power grid.","The top-10 nodes chosen by HGC produce faster-growing SIR infection curves than baseline top-10 sets in Jazz, USAir, NS, and Facebook.","HGC's monotonicity is the highest of all methods on every network studied, so its scores separate nodes more finely and produce fewer tied ranks.","Because the method uses only second-order neighbourhoods ($R=2$) and two propagation iterations ($T=2$), it remains a local computation once cycle ratios are known."],"supporting_citations":[{"why":"Supplies the gravity formula that HGC modifies by changing the distance metric.","marker":"[12]"},{"why":"Defines the local gravity model (LGM), the main gravity baseline HGC extends and is compared against.","marker":"[13]"},{"why":"Introduces effective distance as a probability-based asymmetric distance, the core replacement for topological distance.","marker":"[25]"},{"why":"Provides the network constraint coefficient used to weight structural holes in the gravity term.","marker":"[20]"},{"why":"Defines the cycle number matrix and cycle ratio that seed the RCP propagation term.","marker":"[8]"},{"why":"The SIR model is used as the ground-truth spreading process for evaluating rankings.","marker":"[21]"},{"why":"Kendall's tau is the accuracy metric used to compare ranking lists.","marker":"[22]"},{"why":"Jaccard similarity is used to compare top-k spreader lists.","marker":"[23]"},{"why":"Monotonicity is used to measure how well methods distinguish node ranks.","marker":"[24]"},{"why":"Provides the epidemic threshold formula used to set the SIR infection probability.","marker":"[28]"}],"fun_headline_variants":["Gravity with asymmetric distance and cycles beats 7 spreader-finding methods","New HGC score ranks spreaders better using cycles and directed distances","Cycle structure plus asymmetric distance lifts spreader detection","Asymmetric gravity and cycles: HGC finds influential spreaders better","Gravity model gets cycle boost to identify spreaders, beats 7"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the number of shortest cycles passing through each pair of nodes can be computed for any network, but no algorithm or complexity bound is given for that computation, so the method's rankings on large networks like Facebook and Power depend on an unspecified cycle-enumeration choice.","fun_headline_variants_meta":{"raw":{"variants":["Gravity with asymmetric distance and cycles beats 7 spreader-finding methods","New HGC score ranks spreaders better using cycles and directed distances","Cycle structure plus asymmetric distance lifts spreader detection","Asymmetric gravity and cycles: HGC finds influential spreaders better","Gravity model gets cycle boost to identify spreaders, beats 7"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2136,"prompt_tokens":1040,"completion_tokens":1096,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1006}},"tokens_in":656,"tokens_out":1096,"duration_ms":8440,"temperature":1.0,"reasoning_tokens":1006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:27:59.677424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a tree network (no cycles at all): HGC's RCP term is zero for every node, so HGC collapses to its gravity term; if HGC then still outperforms LGM by the same margin it shows on cyclic networks, the cycle component is not the source of the claimed advantage. Alternatively, rerun the Facebook comparison with a different shortest-cycle enumeration and check whether the top-k list changes materially.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gravity formula that HGC modifies by changing the distance metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the local gravity model (LGM), the main gravity baseline HGC extends and is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the cycle number matrix and cycle ratio that seed the RCP propagation term."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Monotonicity is used to measure how well methods distinguish node ranks."}],"review_version":1}