{"id":"98e8b629-3c7f-467c-acff-d6051964ebd1","arxiv_id":"2506.18641","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A subgraph-extraction method combining degree-based node removal with average-degree-matching edge pruning preserves SIR spreading dynamics and information flow in heterogeneous networks at over 85% size reduction.","lead":"This paper proposes a fast network reduction method that removes low-centrality nodes and then prunes edges to keep the smaller network's average degree close to the original, and shows that it preserves epidemic-spreading and information-flow dynamics on scale-free and real-world networks. A smart generalist should read it because it offers a practical way to simulate disease or information spread on much smaller network stand-ins with similar outcomes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim lacks an ablation: edge pruning enforces ⟨k⟩≈⟨k⟩₀, yet no test isolates whether DC+-ordered node removal contributes beyond any node removal followed by the same pruning.","rationale":"The paper's central assertion is that the combined NRDC′+ strategy, DC+-based node removal followed by average-degree-matching edge pruning, accurately preserves SIR spreading dynamics and information flow on heterogeneous networks while reducing size by more than 85%. For this assertion to be meaningful, both algorithmic components must contribute, or at least the node-ordering component must be shown to be necessary. The reported experiments establish that adding edge pruning to NRDC+ dramatically improves foverlap (Table 2), and that NRDC′+ usually beats four sampling baselines. However, the baselines are not given the same average-degree-matching post-processing, so the comparison does not isolate the value of the DC+ ordering. This is a load-bearing gap because the paper's novelty claim centers on 'coordinated optimization of node removal and edge pruning,' and its conclusion attributes redundant roles specifically to low-DC+ nodes. A randomized ablation with identical pruning would settle whether the ordering matters or whether the average-degree-matching pruning is doing essentially all of the work. The Reader's weakest-assumption concern about kmin tuning is justified but not fatal to the headline numbers, since Table 2 with fixed kmin = 2 still yields foverlap values above 0.9 for all twelve networks; the missing ablation is more fundamental to the causal claim. The paper has strengths: consistent results across many real networks, a simple and fast algorithm, and comparison against several renormalization and sampling methods. It also has no formal verification and no released code, which raises the value of a concrete ablation check. Since the reported evidence is compatible with the claim but does not yet establish the centrality-ordering contribution, the existing CONDITIONAL verdict remains appropriate.","tokens_in":34342,"tokens_out":5026,"duration_ms":64836,"concrete_test":"Run a three-arm ablation on the twelve real networks: (1) remove the same fraction q of nodes selected uniformly at random, (2) remove nodes by plain degree centrality, and (3) remove nodes by DC+, then apply Algorithm 1 unchanged (kmin = 2, δ = 0.01) to each residual network. Recompute the Table 2 foverlap values for l = 3. If the random and plain-degree variants remain within roughly 0.02 of the NRDC′+ column, the DC+ ordering is negligible and the claim should be re-scoped; if they fall below 0.9 for several networks, the centrality ordering is essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim bundles two algorithmic decisions: DC+-ordered node removal (§3, step 1) and average-degree-matching edge pruning (§3, Algorithm 1). Table 2 isolates the value of pruning only by comparing NRDC+ (node removal only) with NRDC′+ (node removal plus pruning) for each network: foverlap jumps from roughly 0.62–0.83 to 0.89–0.99. The paper never runs the complementary ablation: random or plain-degree node removal followed by the same edge-pruning algorithm. Because Algorithm 1 explicitly drives the subgraph average degree to within δ = 0.01 of ⟨k⟩₀, and because the SIR final-size curve and the small-τ expansion of the normalized partition function, Zτ/N = 1 − ⟨k⟩τ + O(τ²), both depend on average degree at leading order, the entire performance gap could be produced by the pruning step alone. If that is the case, the 'enhanced degree centrality' component is not load-bearing, and the conclusion that low-DC+ nodes are redundant in information transmission is not established by the reported experiments. The central claim is not disproven, but the causal attribution in the abstract and Section 5 is underdetermined. The Reader's kmin concern is real but less decisive here, since Table 2 already exceeds 0.9 on all twelve networks with the fixed kmin = 2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage subgraph extraction method for complex networks. Stage 1 removes nodes in ascending order of an 'enhanced degree centrality' DC+ (degree × average neighbor degree) up to a preset removal ratio q. Stage 2 applies an edge-pruning algorithm (Algorithm 1) that deletes edges to low-degree neighbors while preserving connectivity, with the explicit objective of bringing the subgraph's average degree to within δ = 0.01 of the original average degree. The method, called NRDC′+, is evaluated on ER random graphs, BA scale-free networks, and twelve real-world networks by comparing SIR epidemic curves (via an foverlap measure) and normalized Laplacian partition functions Zτ/N between the original graph and the reduced subgraphs. The authors report that subgraphs retaining 1/8 (and in some cases 1/32) of the nodes preserve the dynamics with foverlap mostly above 0.9, and that NRDC′+ usually outperforms four sampling baselines and three renormalization methods.","tokens_in":34570,"tokens_out":6737,"duration_ms":72371,"significance":"The result is potentially valuable if it holds: it offers a computationally cheap way to obtain a small subgraph that reproduces SIR epidemic saturation curves and the normalized Laplacian partition function of much larger heterogeneous networks, with possible applications to epidemic forecasting and social-media intervention. The paper's strengths are its breadth (two synthetic families and twelve real networks spanning seven domains), the use of 100-run averages, a transparent overlap metric, and the comparison with four sampling baselines and three renormalization methods. The computational-speed claims are supported by the timing comparisons in Figure E11. However, the causal attribution of the method's success to the DC+ ordering is not yet supported because the edge-pruning stage is designed to enforce the average degree of the original network and no ablation isolates the ordering decision; moreover, the paper provides no theoretical argument for why low-DC+ nodes should be redundant. These issues are correctable with additional experiments rather than fundamental errors.","major_comments":[{"comment":"The experiments do not isolate the contribution of DC+-ordered node removal. Table 2 shows that foverlap improves from roughly 0.63–0.83 under NRDC+ to roughly 0.91–0.97 under NRDC′+ once edge pruning is added, but no condition applies Algorithm 1 to subgraphs produced by random or plain-degree node removal. Since Algorithm 1 explicitly targets ⟨k⟩≈⟨k⟩0 with δ = 0.01, and since the SIR saturation curve and the leading-order small-τ expansion of the normalized partition function in Eq. (3) both depend on the average degree, the observed preservation could be produced by the pruning stage alone. Please add this ablation and report foverlap for random and plain-degree node removal followed by the same pruning algorithm; if those variants attain comparable foverlap, the conclusion that low-DC+ nodes are redundant is not supported by the current evidence.","section":"Section 3 (Algorithm 1) and Table 2"},{"comment":"The paper states that the optimal kmin varies by network (near 2 for Music, near 10 for Metabolic) yet fixes kmin = 2 for Table 2; the comparisons against renormalization methods in Figures E3–E10 use kmin = 3 for Blogs, kmin = 13 for Metabolic and Drosophila, and kmin = 2 for Music. The headline comparisons against RG methods are therefore run under network-specific tuning, while the abstract presents the method as a single algorithm. Please state clearly which claims use the fixed kmin = 2 configuration, provide a sensitivity analysis of foverlap versus kmin across all twelve networks, and reconcile the reported optimal value of about 10 for Metabolic with the value 13 used in Figures E4 and E8.","section":"Section 4.4 (\"The setting of parameter kmin\") and Figures E3–E10"},{"comment":"The final conclusion that 'nodes and edges discarded during the subgraph extraction process play a redundant role in information transmission' goes beyond the evidence. The experiments show aggregate curve overlap for ρr and Zτ, not that individual low-DC+ nodes or pruned edges carry no information. Because the pruning step restores the average degree by construction, the redundancy interpretation depends on the missing ablation identified above. Until that ablation is provided, this statement should be rephrased as a hypothesis or restricted to the observed aggregate-level preservation.","section":"Section 5 and Abstract"}],"minor_comments":[{"comment":"The phrase 'almost always achieves the highest accuracy' is not quantified. For l = 3, some baseline methods beat NRDC′+ on specific networks (e.g., CNARW 0.9834 vs. 0.9489 on Metabolic; MHRW 0.9530 vs. 0.9142 on Drosophila; MHRW 0.9391 vs. 0.9210 on USpowergrid). Please report the win rate and discuss these exceptions explicitly.","section":"Table 2 and Section 4.5"},{"comment":"The labels 'RNDC+' and 'RNDC′+' should be 'NRDC+' and 'NRDC′+' for consistency with the main text.","section":"Figures E1–E10"},{"comment":"The caption states 'the initial Internet network' but the figure shows the Metabolic network; please correct this mismatch.","section":"Figure 6 caption"},{"comment":"The name 'Uspowergrid' appears with inconsistent capitalization; use 'USpowergrid' consistently.","section":"Section 4.2 and Table 1"},{"comment":"The SIR dynamics are simulated on the network with EoN, but Eqs. (1) and (2) are written as mass-action ODEs; please clarify whether these equations are only illustrative or are actually used to generate the reported curves.","section":"Section 4.3"},{"comment":"The paper provides URLs for datasets and baseline implementations but not for the authors' own NRDC′+ implementation; please release the code to support reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a network-dynamics venues, and the edge-pruning component is a useful addition to the authors' earlier centrality-based node-removal work. The decisive issue is the missing ablation separating the effect of DC+-ordered removal from the effect of average-degree-matching pruning. If the ablation shows that random or degree-based removal plus pruning performs comparably, the central attribution in the abstract and conclusion must be revised. The kmin-tuning inconsistency in the RG comparisons should also be reported transparently. I do not see grounds for rejection, but the current evidence is insufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid empirical contribution: it shows you can shrink heterogeneous networks by 85% and still reproduce SIR spreading curves and the normalized partition function, with foverlap mostly above 0.9 on 12 real networks. The combination of DC+-based node removal with an edge-pruning step that enforces original average degree is new, and the validation against four sampling and three renormalization baselines is extensive. The foverlap metric is sensible, and the authors are honest that spreading ability correlates with average degree.\n\nThe soft spots are real, though not fatal. The biggest issue is the missing ablation. Table 2 isolates the pruning contribution by comparing NRDC+ with NRDC'+, but no one runs random or plain-degree node removal followed by the same pruning. Since Algorithm 1 forces the subgraph average degree to within 0.01 of the original, and both the SIR final size and the small-tau expansion of Z_tau/N depend on average degree at leading order, the pruning step alone could explain the performance jump. The DC+ ordering may be doing very little work, so the paper's causal claim about low-DC+ nodes being redundant is underdetermined.\n\nThe kmin concern is real but less decisive: kmin=2 works across all twelve networks in Table 2, even though the authors admit the optimal value is network-dependent (Metabolic near 10, and kmin=13 in the renormalization comparison). The generality claim is somewhat overstated, but the fixed parameter still delivers. Missing error bars and no released code are minor issues.\n\nOverall, the engineering result is useful for practitioners who need a fast way to shrink a network for SIR-type simulations or spectral estimates. The conceptual conclusion about redundancy of low-DC+ nodes does not follow from the experiments as reported. I would send this to a serious referee, with a clear request for the ablation and a sensitivity analysis on kmin.","headline":"Useful empirical network-reduction method, but the claim that low-DC+ nodes are redundant is undercut by a missing ablation—pruning alone likely explains the preserved dynamics.","tokens_in":35121,"tokens_out":2556,"would_cite":true,"duration_ms":29861,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C82","68R10"],"pacs":["89.75.Hc","89.75.Fb"],"model":"deepseek-v4-flash","headline":"This paper claims that removing low-centrality nodes and then pruning edges to match average degree yields a subgraph one-eighth the size that still reproduces SIR epidemic spreading and the Laplacian partition function.","keywords":["network reduction","subgraph extraction","enhanced degree centrality","SIR spreading dynamics","information flow","partition function","edge pruning","complex networks"],"falsifier":"Run the NRDC'+ method on the Metabolic network at $k_{\\min}=2$ and at $k_{\\min}=13$ and compare the SIR overlap $f_{\\mathrm{overlap}}$: if the $k_{\\min}=2$ subgraph scores below $0.9$ while the $k_{\\min}=13$ subgraph scores near $0.95$, the claimed network-independent $k_{\\min}=2$ setting fails. A held-out test on another scale-free network, locating the optimal $k_{\\min}$ by a mean-absolute-error scan, would settle whether $k_{\\min}=2$ lies on a flat plateau or must be tuned per network.","tokens_in":34092,"feed_emoji":"🦠","tokens_out":8599,"duration_ms":87484,"temperature":0.7,"pith_summary":"This paper claims that a two-step subgraph extraction method can shrink a heterogeneous real-world network to $1/8$ or even $1/32$ of its nodes while still reproducing the original's SIR epidemic spreading curves and information flow. The first step removes nodes with the smallest enhanced degree centrality, the product of a node's degree and its average neighbor degree, which tends to raise the subgraph's average degree; the second step prunes edges until the average degree matches the original network's. On twelve real networks and the Barabási-Albert scale-free model, the pruned subgraphs achieve overlap scores above $0.9$ and almost always beat four sampling baselines, while Erdős-Rényi random graphs do not show this self-similarity. If correct, large-scale epidemic and information-diffusion simulations could be run on small surrogate networks at much lower cost, and many discarded low-centrality nodes and edges would be shown to be dynamically redundant.","feed_headline":"Shrink a network to one-eighth and keep its spreading dynamics","feed_subtitle":"Two-step removal and pruning preserves SIR epidemics and information flow, beating four sampling baselines.","key_machinery":"Two objects carry the argument. First, enhanced degree centrality $\\mathrm{DC}^+(i)=k_i\\times\\bar{k}_{\\mathrm{nn}}(i)$, the product of a node's degree and its average neighbor degree, orders node deletion: nodes are removed in ascending $\\mathrm{DC}^+$ order, preferentially discarding low-centrality nodes while retaining hubs. Second, Algorithm 1 prunes edges: it repeatedly selects a node of degree above $k_{\\min}$, removes the edge to its lowest-degree neighbor, restores the edge if connectivity breaks, and stops when the subgraph's average degree falls within $\\delta=0.01$ of the original. The measured quantities are the SIR curves $r(t)$ and $i(t)$, the final-size curve $\\rho_r(\\beta)$, the normalized partition function $\\bar{Z}_{\\tau,l}=Z_{\\tau,l}/N_l$ with $Z_{\\tau,l}=\\mathrm{Tr}(e^{-\\tau L_l})$, and the overlap score $f_{\\mathrm{overlap}}=1/(1+S_\\Delta)$ comparing $\\rho_r$ curves via Simpson integration.","core_discovery":"The central discovery is that preserving the average degree during node-removal-based reduction is enough to preserve both SIR spreading dynamics and the normalized Laplacian partition function $\\bar{Z}_{\\tau,l}=Z_{\\tau,l}/N_l$ in heterogeneous networks, provided node removal is ordered by the enhanced degree centrality $\\mathrm{DC}^+$. Merely removing low-$\\mathrm{DC}^+$ nodes makes subnetworks spread faster than the original because their average degree rises; adding the edge-pruning step restores the average degree and brings the spreading curves and partition-function curves back onto the original network's. The paper reports $f_{\\mathrm{overlap}}$ values computed from the $\\beta$-dependence of the final epidemic size $\\rho_r$, with most real networks above $0.9$ at $l=3$ ($1/8$ of nodes) and Enron still near $0.98$ at $l=5$ ($1/32$ of nodes). Because the same pruned subgraphs also reproduce $Z_\\tau$, the method simultaneously preserves macroscopic information-flow properties, not just one epidemic observable.","pith_inferences":["Editorial inference: if average-degree matching is the operative mechanism, the method effectively pins the first moment of the degree distribution, and a testable prediction is that any two reductions with the same average degree will show similar early-time SIR growth even if their degree correlations differ.","Editorial inference: the optimal $k_{\\min}$ variation found by the authors, near 2 for Music, near 10 for Metabolic, and 13 used for two networks in the renormalization comparison, suggests that a fully parameter-free version should select $k_{\\min}$ adaptively; the reported $k_{\\min}=2$ results may understate the method's ceiling on some networks.","Editorial inference: the connectome result invites a direct transfer test, applying the same reduction to other empirically multiscale systems such as parcellated brain atlases or layered transportation networks, to see whether the observed self-similarity is a general property of hierarchical real-world networks.","Editorial inference: because $\\mathrm{DC}^+$ depends on neighbor degrees, the removal order encodes degree assortativity, so on strongly disassortative networks low-$\\mathrm{DC}^+$ nodes might be structural bridges; the connectivity-restoring pruning step would then be the load-bearing component, a behavior the current experiments do not isolate."],"forward_implications":["On twelve real heterogeneous networks and Barabási-Albert scale-free networks, subgraphs containing only $1/8$ of the nodes reproduce the original SIR spreading curves with $f_{\\mathrm{overlap}} > 0.9$, so epidemic simulations can be run on the small subgraph instead of the full network.","For the largest tested networks, Internet and Enron, the method stays accurate at $1/24$ and $1/32$ scales, so the reduction does not saturate at moderate compression.","Because the partition function $Z_\\tau$ is also preserved, the same reduced subgraph can stand in for the full network when computing spectral entropy and free-energy estimates of information diffusion.","The human connectome application shows that reduction ratios $q=0.54$ to $0.92$ reproduce empirical multiscale spreading curves, suggesting the method transfers to hierarchically organized biological networks.","The $O(N_0+M_0)$ node-removal complexity makes the method several orders of magnitude faster than spectral renormalization approaches, so it scales to networks with millions of nodes."],"supporting_citations":[{"why":"Defines enhanced degree centrality DC+, the product of degree and average neighbor degree, which orders the node removal.","marker":"[5]"},{"why":"Supplies the Erdős-Rényi random graph model used as a synthetic test network.","marker":"[42]"},{"why":"Supplies the Barabási-Albert scale-free model used as a synthetic test network.","marker":"[43]"},{"why":"Defines the SIR compartmental model and simulation parameters used as the spreading-dynamics observable.","marker":"[44]"},{"why":"Introduces spectral entropy and the network density matrix used to define information flow.","marker":"[45]"},{"why":"Introduces the partition function $Z_\\tau=\\mathrm{Tr}(e^{-\\tau L})$ as the macroscopic measure of information flow.","marker":"[46]"},{"why":"Provides the coarse-graining-via-partition-function baseline and the information-flow framing this method matches.","marker":"[26]"},{"why":"Supplies the RDN graph-sampling baseline and the sampling comparison methodology.","marker":"[29]"},{"why":"Supplies the Laplacian renormalization group baseline the method is compared against.","marker":"[20]"},{"why":"Supplies the geometric renormalization baseline used in the renormalization comparisons.","marker":"[14]"}],"fun_headline_variants":["Two-step prune keeps SIR spread and info flow in 1/8 network","Cut node count by 87%, keep epidemic curves and info flow","One-eighth network, full SIR and info flow preserved","Prune edges to match degree: keep dynamics at 1/8 size","Reducing networks by 85% without losing SIR or info flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single edge-pruning cutoff, the smallest node degree allowed to lose edges, set to $k_{\\min}=2$, works across all real networks, even though the paper's own scan finds optimal values near 10 for the Metabolic network and uses $k_{\\min}=13$ for Metabolic and Drosophila in the renormalization comparisons.","fun_headline_variants_meta":{"raw":{"variants":["Two-step prune keeps SIR spread and info flow in 1/8 network","Cut node count by 87%, keep epidemic curves and info flow","One-eighth network, full SIR and info flow preserved","Prune edges to match degree: keep dynamics at 1/8 size","Reducing networks by 85% without losing SIR or info flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000819,"raw_usage":{"total_tokens":3614,"prompt_tokens":1002,"completion_tokens":2612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2516}},"tokens_in":618,"tokens_out":2612,"duration_ms":20906,"temperature":1.0,"reasoning_tokens":2516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:15:30.711616+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the NRDC'+ method on the Metabolic network at $k_{\\min}=2$ and at $k_{\\min}=13$ and compare the SIR overlap $f_{\\mathrm{overlap}}$: if the $k_{\\min}=2$ subgraph scores below $0.9$ while the $k_{\\min}=13$ subgraph scores near $0.95$, the claimed network-independent $k_{\\min}=2$ setting fails. A held-out test on another scale-free network, locating the optimal $k_{\\min}$ by a mean-absolute-error scan, would settle whether $k_{\\min}=2$ lies on a flat plateau or must be tuned per network.","supporting_citations":[{"cited_title":"Identification of influential nodes in complex networks with degree and average neighbor degree","cited_arxiv_id":null,"evidence_quote":"Defines enhanced degree centrality DC+, the product of degree and average neighbor degree, which orders the node removal."},{"cited_title":"Erdös and A","cited_arxiv_id":null,"evidence_quote":"Supplies the Erdős-Rényi random graph model used as a synthetic test network."},{"cited_title":"Emergence of scaling in random networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Barabási-Albert scale-free model used as a synthetic test network."},{"cited_title":"Epidemic processes in complex networks","cited_arxiv_id":null,"evidence_quote":"Defines the SIR compartmental model and simulation parameters used as the spreading-dynamics observable."},{"cited_title":"Spectral entropies as information-theoretic tools for complex network comparison","cited_arxiv_id":null,"evidence_quote":"Introduces spectral entropy and the network density matrix used to define information flow."},{"cited_title":"Diversity of information pathways drives sparsity in real-world networks","cited_arxiv_id":null,"evidence_quote":"Introduces the partition function $Z_\\tau=\\mathrm{Tr}(e^{-\\tau L})$ as the macroscopic measure of information flow."},{"cited_title":"Coarse-graining network flow through statistical physics and machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the coarse-graining-via-partition-function baseline and the information-flow framing this method matches."},{"cited_title":"Sampling from large graphs","cited_arxiv_id":null,"evidence_quote":"Supplies the RDN graph-sampling baseline and the sampling comparison methodology."},{"cited_title":"Laplacian renormalization group for heterogeneous networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Laplacian renormalization group baseline the method is compared against."},{"cited_title":"Ángeles Serrano","cited_arxiv_id":null,"evidence_quote":"Supplies the geometric renormalization baseline used in the renormalization comparisons."}],"review_version":1}