{"id":"659642d0-b79b-4b4b-9110-ff3dfee10c93","arxiv_id":"2502.03662","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"EC-SBM is a scalable synthetic network generator that preserves cluster edge connectivity and degree sequence better than SBM or RECCS.","lead":"A new program, EC-SBM, creates fake networks that copy the shape and cluster structure of real networks. It is meant to give researchers better test cases for community detection algorithms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported accuracy comparison is vulnerable to selection bias: the input clusterings compared in Experiment 2 were chosen using the same 74 test networks in Experiment 1.","rationale":"The reader's conditional verdict is appropriate, but the decisive weakness is located differently. The Section 5 limitation about not studying clustering algorithms is honest and narrows the stated contribution; it does not by itself falsify the fidelity claim. The stronger problem is that the empirical comparison is circular in a statistical sense: the same data are used to choose and to test the winning configuration. The authors select SBM+WCC and Leiden-Mod+CM based on the same 74 networks on which the final comparisons are drawn, and no significance testing is provided. This makes the central empirical claim fragile. The proposed holdout test is straightforward because the software and corpus are public. If the holdout test supports the ranking, the paper should be acceptable with the disclosed limitations; if not, the abstract's 'generally more accurate' claim should be softened.","tokens_in":15331,"tokens_out":8894,"duration_ms":86769,"concrete_test":"Perform a nested holdout evaluation on the same 74 networks: randomly split into 50 selection and 24 evaluation networks (or 5-fold cross-validation). For each split, use only the selection networks to pick each generator's best input clustering from the 10 candidates (as in Experiment 1), then compute all eight statistics on the held-out evaluation networks for EC-SBM, RECCS, and SBM with their selected clusterings. Report median distances with bootstrap 95% confidence intervals or paired sign tests over networks. If EC-SBM's advantage over RECCS/SBM persists on held-out networks, the selection-bias concern is resolved; if it shrinks or reverses, the headline claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the protocol pairing Experiments 1 and 2 (Section 4.1). In Experiment 1, the authors visually inspect distance plots on all 74 small/medium networks to select SBM+WCC and Leiden-Mod+CM as the 'best' input clusterings for SBM, RECCS, and EC-SBM. Experiment 2 then compares the three generators on those same 74 networks using exactly those selected clusterings (Figs. 6-7). This is selection-on-test-data: the configuration that wins the comparison was chosen by looking at the outcome on the same corpus. No holdout, cross-validation, or multiple-testing correction is reported, and Figs. 6-7 show only medians with no error bars or significance tests. Under this protocol, even a generator with no real advantage could appear 'generally more accurate' by capitalizing on noise in the 74-network corpus. Because the abstract's central claim is an empirical superiority claim ('generally more accurate than current alternative approaches'), this bias directly threatens the headline. The proxy concern raised in Section 5 is real but explicitly disclosed as future work; the selection bias is not disclosed and is internal to the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EC-SBM, a three-stage synthetic network generator that takes a clustered real-world network as input and produces a synthetic network with the same cluster assignment. Stage 1 constructs for each cluster a spanning subgraph that is guaranteed to have edge connectivity at least the empirical value, Stage 2 generates an outlier subnetwork via SBM, and Stage 3 adds edges to match the input degree sequence. The paper proves (Theorem 1) that the spanning-subgraph construction is k-edge-connected, and it reports an empirical comparison with SBM and RECCS on 74 networks, plus runtime results on three large networks, concluding that EC-SBM is generally more accurate than the alternatives and scales to millions of nodes.","tokens_in":15500,"tokens_out":11431,"duration_ms":100153,"significance":"If the empirical claims hold, EC-SBM would be a useful, scalable tool for generating synthetic networks with ground-truth clusters that preserve cluster edge connectivity and degree sequence, addressing a known weakness of SBM-based generators. The paper contributes a public code repository, a clear theorem with a sound proof for the core construction, and a substantial corpus (74 networks) for evaluation. However, the significance is tempered by two methodological issues discussed below: the comparison protocol selects configurations on the same test data used for the final evaluation, and several of the reported accuracy metrics are enforced by construction rather than independently validated. The paper is also candid in Section 5 that it does not test clustering algorithms on the generated networks, which limits the direct support for the stated benchmarking motivation.","major_comments":[{"comment":"The protocol for selecting input clusterings creates a selection-on-test-data problem. Experiment 1 uses the same 74 networks to choose SBM+WCC and Leiden-Mod+CM as the best clusterings (via a qualitative reading of Figure 5), and Experiment 2 then reports the generator comparison on those same 74 networks using exactly these clusterings (Figures 6–7). Because the choice of which clusterings to report was informed by the outcome on this corpus, the comparative claim that EC-SBM is 'generally more accurate' is not supported as a general statement; it is at best a description of this corpus under a post-hoc selected configuration. The paper reports only medians in Figures 6–7 and provides no error bars, confidence intervals, or significance tests, so the magnitude and reliability of the differences are unclear. A valid protocol would either select clusterings on a training subset and evaluate on a held-out subset, or report results across all ten clusterings with appropriate multiple-testing awareness.","section":"§4.1–4.3"},{"comment":"The evaluation treats degree sequence, outlier degree, and mincuts as measures of accuracy, but these are partly enforced by construction. Stage 1 (Section 6.2.2) explicitly constructs a spanning subgraph that is λ_i-edge-connected for each cluster, and Stage 3 adds edges to match the input degree sequence. Consequently, the near-zero RMSE values for degree and outlier degree, and the lower-bound direction of mincuts in Figure 7, are direct consequences of the algorithm's design rather than independent evidence of fidelity. The manuscript should state this explicitly and base its accuracy claims primarily on statistics that are not directly imposed—such as pseudo-diameter, characteristic time, global clustering coefficient, and mixing parameter—or otherwise justify why enforced properties should count as validation.","section":"§3, §6.2, Table 1, Figure 7"}],"minor_comments":[{"comment":"The sentence 'We setmicro ers and micro degs to true' appears to have a formatting error; it should read 'We set micro_ers and micro_degs to true'.","section":"§2.1"},{"comment":"There is a duplicated passage in the text around Figure 5/6: the sentence beginning 'degree sequence. On the other hand, EC-SBM using SBM+WCC input clustering is a well-rounded pipeline...' appears twice, and the flow is broken.","section":"§4.3"},{"comment":"References [12] and [13] list the same DOI (10.1017/nws.2020.45); the ABCD+o paper should have a distinct DOI.","section":"References"},{"comment":"The boxplots in Figure 5 are difficult to read because of the many categories and small labels; consider splitting into separate panels or increasing the font size.","section":"Figure 5"},{"comment":"The text first says vertices are 'arbitrarily labeled' and later says the implementation processes them in decreasing order of degree; the proof of Theorem 1 holds for any order, but the exposition should clarify that the degree-ordered variant is an implementation choice, not a requirement of the theorem.","section":"§6.2.2"},{"comment":"The column header 'CEN orkut livejournal' lacks separators and is confusing; please format the table with clear per-column headers for each network.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The selection-bias issue is the most serious problem: the empirical superiority claim in the abstract rests on a comparison where the clusterings were chosen using the same corpus on which the final comparison is reported. This needs to be addressed with a proper validation protocol or by clearly reframing the results as corpus-specific. The circularity of the degree and edge-connectivity metrics is also worth raising with the authors. The theorem and code appear sound, so I believe the paper is salvageable with major revisions rather than rejectable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: EC-SBM is a real, incremental improvement over RECCS, with the k-edge-connected skeleton as the genuinely new idea, and the code is public. The theory is fine. The experimental comparison, however, has a protocol problem that makes the \"generally more accurate\" claim weaker than the paper suggests.\n\nWhat's new: the three-stage pipeline—build a k-edge-connected spanning subgraph per cluster, augment with SBM, generate the outlier subnetwork, then degree-correct—is not in RECCS or plain SBM. Theorem 1 is a simple and correct argument that the skeleton is k-edge-connected; I checked the two cases and it holds. The paper also does a useful systematic comparison of input clustering methods (SBM+WCC vs Leiden-Mod+CM etc.), which is practically valuable. Scaling to millions of nodes with runtimes around 1–5 hours is credible for the graph-tool-based approach.\n\nSoft spots, in order of severity.\n\nFirst, selection bias. Experiment 1 examines all 74 networks, picks the \"best\" input clustering per simulator using those very networks, and Experiment 2 then compares the three simulators on the same 74. No holdout, no correction. The abstract says \"generally more accurate\"—that's the central empirical claim, and this protocol can manufacture an advantage. This is the load-bearing flaw. It doesn't invalidate the method, but it does invalidate the unqualified superiority claim.\n\nSecond, several RMSEs are close to zero by construction. Stage 1 forces edge connectivity at least lambda_i and Stage 3 matches degrees; so low RMSE on mincuts, degree sequence, and outlier degree is partly baked in. The paper says this implicitly but doesn't frame it as a limitation. Comparing two simulators on statistics one of them is engineered to hit is not an apples-to-apples test.\n\nThird, the proxy assumption is explicitly acknowledged in Section 5: they don't test whether clustering algorithms perform better on EC-SBM networks. That's fine as a boundary, but it means the motivating claim—better benchmarking—is unverified.\n\nMinor: no error bars or significance tests; medians over 74 networks without variance is thin.\n\nRecommendation: this deserves a serious referee, and the right outcome is conditional acceptance with a request for a corrected comparison (e.g., a held-out set, or at minimum significance testing and a clear statement of which metrics are guaranteed by the algorithm). The core construction is a contribution; the evaluation needs to be honest about what it can and cannot show.","headline":"A useful new synthetic-network generator with a clean theoretical core, but the headline accuracy comparison is undercut by selection on the test corpus.","tokens_in":16051,"tokens_out":1560,"would_cite":true,"duration_ms":14183,"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":"EC-SBM generates synthetic networks that match real communities' internal edge connectivity, with higher fidelity than current SBM-based simulators.","keywords":["synthetic network generation","community detection","stochastic block model","edge connectivity","degree sequence","graph clustering","network simulation","real-world networks"],"falsifier":"Take any network from the 74-network corpus, generate an EC-SBM twin, and rerun the evaluation after deleting the GEN-KECSSN stage; if the eight statistics barely change, the connectivity mechanism is not load-bearing. Separately, run a standard community detection algorithm on both the real and synthetic networks and compare recovery of the planted clusters; if recovery patterns diverge sharply, the eight-statistic match is not sufficient for benchmarking.","tokens_in":15112,"feed_emoji":"🕸️","tokens_out":9370,"duration_ms":83411,"temperature":0.7,"pith_summary":"This paper proposes EC-SBM, a synthetic network generator that takes a clustered real-world network and produces a new network with the same cluster assignment but freshly generated edges. The central claim is that EC-SBM reproduces both network-level statistics (degree sequence, pseudo-diameter, characteristic time, global clustering coefficient) and community-level statistics (per-cluster edge connectivity, internal edge counts, mixing parameters, outlier degrees) more accurately than plain SBM and than the RECCS simulator. The key design choice is to guarantee each synthetic cluster is internally edge-connected up to the level of the original cluster, then fill in remaining edges with an SBM and correct degrees. If the claim holds, community detection researchers get a scalable way to build benchmark graphs whose ground-truth communities are known, connected, and statistically similar to real communities.","feed_headline":"EC-SBM reproduces real community structure at scale","feed_subtitle":"Three-stage SBM-based generator beats prior simulators on community and network statistics at scale.","key_machinery":"The load-bearing object is the GEN-KECSSN procedure that generates a $k$-edge-connected spanning subnetwork on a cluster: it starts with a $(k+1)$-clique on the first $k+1$ vertices and then connects every later vertex to $k$ previously placed vertices. The theorem that every edge cut in the resulting graph has size at least $k$ is what converts the empirical minimum-cut values of the input clusters into a lower bound on synthetic cluster connectivity. Around this core, EC-SBM layers a stochastic block model (SBM) to fill in the remaining intra- and inter-cluster edges, a simplification step that collapses parallel edges and self-loops, a separate SBM pass for the outlier subnetwork, and a final degree-correction pass that adds edges so each vertex reaches its target degree.","core_discovery":"On its own terms, the paper's discovery is that realistic synthetic networks for community detection can be produced by a three-stage pipeline that treats cluster connectivity as a hard constraint rather than an emergent property. In Stage 1, for each cluster EC-SBM builds a spanning subgraph with edge connectivity at least the cluster's empirical minimum cut, using a procedure that starts from a $(k+1)$-clique and attaches each remaining vertex to $k$ earlier vertices; a theorem shows the result is $k$-edge-connected. SBM then adds the remaining edges, parallel edges and self-loops are removed, and a final degree-correction stage restores the target degree sequence. Over a corpus of 74 real networks clustered in several ways, EC-SBM with the SBM+WCC input clustering was the best overall at matching the eight statistics, beating RECCS on degree sequence, mixing parameters, and characteristic time, while being only slightly behind RECCS on cluster edge connectivity. On three networks with roughly 3 to 14 million nodes, EC-SBM completed in about one to five hours.","pith_inferences":["A natural next test is to use EC-SBM networks as benchmark data for community detection and check whether algorithms' relative rankings match those obtained on real networks; the paper leaves this unexplored.","Because the $k$-edge-connected subgraph construction is independent of SBM, the same idea could be grafted onto other generators, such as LFR or ABCD, to enforce connected ground-truth clusters.","The slight trade-off between EC-SBM and RECCS suggests a hybrid that runs EC-SBM and then applies RECCS's final connectivity refinement might dominate both on cluster edge connectivity without losing degree fidelity.","If the eight-statistic proxy is validated, EC-SBM could also serve as a null model for studying when community detection fails: by varying cluster edge connectivity while holding other statistics fixed, one can isolate the role of internal connectivity."],"forward_implications":["Benchmark designers can generate synthetic graphs whose ground-truth communities are connected by construction, removing an artifact of plain SBM where 30 to 70 percent of clusters come out internally disconnected.","The recommended recipe is to cluster the reference network with SBM followed by well-connected-clusters post-processing, since that input clustering gave EC-SBM its best fidelity.","EC-SBM's runtime of hours on million-node networks makes it a feasible default where deep graph generators run out of memory or require excessive training data.","For users whose priority is exact cluster edge connectivity, RECCS still has a slight edge, so the choice between generators depends on whether degree sequence and mixing fidelity or edge connectivity is more important."],"supporting_citations":[{"why":"Supplies the SBM generation routine used in Stages 1 and 2 to fill in edges after the k-edge-connected core is fixed.","marker":"[32]"},{"why":"Defines the RECCS simulator, the main comparison baseline, and the degree-correction step that EC-SBM adopts as Stage 3.","marker":"[33]"},{"why":"The prior systematic assessment that established which network statistics (diameter, characteristic time, clustering coefficient, degree) are used for fidelity comparison.","marker":"[31]"},{"why":"Provides the well-connected-clusters (WCC) post-processing whose clustering gave EC-SBM and SBM their best results in Experiment 1.","marker":"[37]"},{"why":"Provides the connectivity-modifier (CM) post-processing used to generate alternative input clusterings for comparison.","marker":"[36]"},{"why":"Formalizes the stochastic block model, the base generative model on which EC-SBM's SBM steps depend.","marker":"[23]"}],"fun_headline_variants":["EC-SBM: synthetic networks that mirror real communities","EC-SBM preserves cluster ties for realistic synthetic networks","EC-SBM enforces cluster edge connectivity for realism","Scale to millions: EC-SBM matches real communities accurately"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that matching the eight measured statistics, including edge connectivity and degree sequence, makes a synthetic network a valid stand-in for a real network when benchmarking community detection; the paper does not itself test any clustering algorithm on its generated networks.","fun_headline_variants_meta":{"raw":{"variants":["EC-SBM: synthetic networks that mirror real communities","EC-SBM preserves cluster ties for realistic synthetic networks","EC-SBM enforces cluster edge connectivity for realism","Scale to millions: EC-SBM matches real communities accurately"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2748,"prompt_tokens":877,"completion_tokens":1871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":1807}},"tokens_in":493,"tokens_out":1871,"duration_ms":13759,"temperature":1.0,"reasoning_tokens":1807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:11:57.268731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any network from the 74-network corpus, generate an EC-SBM twin, and rerun the evaluation after deleting the GEN-KECSSN stage; if the eight statistics barely change, the connectivity mechanism is not load-bearing. Separately, run a standard community detection algorithm on both the real and synthetic networks and compare recovery of the planted clusters; if recovery patterns diverge sharply, the eight-statistic match is not sufficient for benchmarking.","supporting_citations":[{"cited_title":"RECCS: Realistic Cluster Connectivity Simulator for Synthetic Network Generation","cited_arxiv_id":"2502.02050","evidence_quote":"Defines the RECCS simulator, the main comparison baseline, and the degree-correction step that EC-SBM adopts as Stage 3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior systematic assessment that established which network statistics (diameter, characteristic time, clustering coefficient, degree) are used for fidelity comparison."},{"cited_title":"Improved Community Detection using Stochastic Block Models","cited_arxiv_id":"2408.10464","evidence_quote":"Provides the well-connected-clusters (WCC) post-processing whose clustering gave EC-SBM and SBM their best results in Experiment 1."},{"cited_title":"PLOS Complex Systems 1(3), 1–25 (2024) https: //doi.org/10.1371/journal.pcsy.0000009","cited_arxiv_id":null,"evidence_quote":"Provides the connectivity-modifier (CM) post-processing used to generate alternative input clusterings for comparison."}],"review_version":1}