{"id":"5028af42-0283-446d-b7b5-0d29c40bdd80","arxiv_id":"1908.02520","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A projected-bipartite modularity QP and a degree-exponent heuristic for choosing a community detection approach on bipartite networks.","lead":"This paper introduces a new modularity measure for projected bipartite networks, where the expected number of shared links is computed from the original bipartite graph rather than the projection. It compares five community-detection algorithms on synthetic and real networks and proposes a simple rule for choosing between bipartite and projected approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QP heuristic is benchmarked on a configuration-model generative process that matches QP's own null model; real networks with degree correlations, nestedness, or unequal community sizes, plus exponent estimates made by eye, may break the µ2-vs-µ1 rule.","rationale":"The reader's weakest assumption and my load-bearing concern coincide: the heuristic is derived from a synthetic benchmark whose generative mechanism matches QP's configuration-model null, and the real-data exponent estimates are informal. This matters because the paper's central claim is not merely that QP is a coherent modularity, but that it is more appropriate in practice, with a concrete rule for when to project. If the benchmark is not representative of real bipartite structure, the rule could misclassify networks and the practical advantage of QP is unsupported. I do not see an internal mathematical error: the derivation of QP from a rewired bipartite null model is consistent, and the Louvain gain formula in Eq. 9 follows from Eq. 8 once the singleton community's negative null contribution is accounted for. The lack of code and data is a reproducibility concern but not a logical flaw in the argument. The overclaim about Dual Projection is real but secondary, since the headline heuristic already conditions on constraints. Given these considerations, the reader's CONDITIONAL verdict remains appropriate; the concern argues for caution, not rejection.","tokens_in":12609,"tokens_out":15020,"duration_ms":168040,"concrete_test":"Re-run the Section 6.2 synthetic comparison using the same exponent pairs as Figure 9, but generate networks with unequal community sizes (e.g., 100, 250, and 650 bottom nodes) and with degree-community correlations (e.g., high-degree bottom nodes preferentially attached inside their own community, or a nested core-periphery structure). Record the homogeneity/completeness ordering of Standard versus Projected as a function of µ2 - µ1. In parallel, re-estimate µ1 and µ2 for the four KONECT datasets using a maximum-likelihood power-law fit with a goodness-of-fit test. If either the synthetic ordering reverses or any dataset changes its µ1-vs-µ2 classification, the Section 8 rule as stated is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim is the Section 8 heuristic: if µ2 > µ1 use the bipartite graph; if µ2 <= µ1 project and use QP. This rule is derived entirely from synthetic graphs generated in Section 3 by planting C equal-size communities and rewiring a fraction p of stubs uniformly at random. That generative process is essentially the same configuration-model null used to define QP in Eq. 7: fixed bottom degrees q_i, fixed top degrees d_m, stubs joined uniformly. The benchmark is therefore in-sample for QP's assumptions and does not test the central claim under degree correlations, nestedness, or heterogeneous community sizes. The real-data support is weaker: in Section 7 the exponents µ1 and µ2 are read off Figure 11 visually ('roughly map to...'), with no stated fitting procedure, and Table 1 shows Standard and Projected partitions are nearly identical on three of the four datasets. If real bipartite networks violate the configuration-model assumption, or if a maximum-likelihood fit changes the ordering of µ1 and µ2, the recommendation to switch algorithms at µ2 = µ1 has no demonstrated basis. Because the paper's stated goal is practical guidance, this is the most load-bearing unsupported step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies community detection in bipartite networks via modularity maximisation, focusing on how one-mode projection interacts with community structure. The author defines a projected modularity QP (Eq. 7) whose null model rewires the original bipartite graph and then projects, derives the corresponding Louvain gain formula (Eq. 9), and proves that forming an induced graph and projecting commute (Eqs. 10-12). Five algorithms are compared on synthetic bipartite graphs with planted communities and either Poisson or power-law degree distributions, using homogeneity, completeness, and V-measure; the same algorithms are applied to four real networks. The main practical output is a heuristic: use the Dual Projection method, or, when that is unavailable and homogeneity is the priority, use bipartite methods when the top-node degree exponent µ2 exceeds the bottom-node exponent µ1 and projected methods when µ2 ≤ µ1.","tokens_in":12834,"tokens_out":14723,"duration_ms":147532,"significance":"The derivation of QP is self-contained, and the Louvain gain formula and commutation result are mathematically sound; the experimental protocol is careful, using 100 graph realisations, 10 Louvain runs per graph, bootstrap confidence intervals, and three complementary metrics. If the proposed heuristic is valid, it provides practitioners with a simple and actionable rule, and the comparative findings—in particular that detected partitions tend to be more homogeneous than complete—are informative. However, the practical recommendation rests on synthetic data generated by the same configuration-model family that underlies QP's null model, and the real-data analysis offers no ground truth or quantitative exponent estimates; moreover, the experiments show that Q- and QP-based methods often give very similar partitions. The significance is therefore conditional on additional validation outside the configuration-model family.","major_comments":[{"comment":"The heuristic 'µ2 > µ1 use the bipartite graph; µ2 ≤ µ1 project and use QP' is derived from the synthetic benchmark described in Section 3, whose generative process—fixed degree sequences with stubs joined at random within and across communities—is the same configuration-model null used to define QP in Eqs. (6) and (7). The experiments use only equal-size planted communities (250 nodes per community, as stated for the Poisson case) and, as far as the text indicates, power-law exponents in a narrow range; they do not include degree correlations, nestedness, or heterogeneous community sizes. The rule is therefore not demonstrated for networks that violate these configuration-model assumptions, and the practical conclusion in Section 8 is stronger than the evidence. I ask for either additional experiments with these structural variations or a substantially more conditional statement of the heuristic.","section":"§6.2 and §8"},{"comment":"The exponents µ1 and µ2 for the four real datasets are estimated by visual inspection ('roughly map to'), with no fitting procedure, confidence intervals, or sensitivity analysis. Because the heuristic switches regimes at µ2 = µ1, an incorrect ordering from a plausible alternative fit could reverse the recommendation. In addition, Table 1 reports only pairwise similarities between the detected partitions and QP values; there is no ground truth for these networks, so the real-data section does not actually validate the heuristic. A quantitative degree-exponent fit with uncertainty, and a discussion of how sensitive the recommendation is to the estimated ordering, are needed before the practical advice can be accepted.","section":"§7 and Fig. 11"},{"comment":"The paper claims that QP is 'more appropriate' for projected bipartite networks than standard modularity, but the experiments in Fig. 6 and Table 1 show that optimising QP (Projected) and optimising standard Q (Standard) produce nearly identical partitions: they overlap completely in the Poisson case and have V-measures between 0.87 and 1.00 on the real datasets. The text should explicitly characterise the conditions, if any, under which QP and Q lead to materially different partitions. Otherwise the claim should be moderated to state that QP is a principled modularity for projections which, in the settings tested, yields partitions very similar to those found by standard modularity.","section":"§2.1 and Conclusions"}],"minor_comments":[{"comment":"The sentence 'H is reduced whenever two nodes from the same target community are assigned by the algorithms to different communities' is incorrect: splitting a target class across clusters reduces completeness C, not homogeneity H. H is reduced when a detected cluster mixes nodes from different target classes. The subsequent interpretation in the results (H > C indicates splitting) is consistent with the correct definitions, so this is a local correction.","section":"§5"},{"comment":"The text describes the null term as 'the probability of having a link between i and j' in the projected randomly rewired network, but the formula is the expected proportion of edge weight in the projected configuration model; the expected number of common neighbours is q_i q_j (∑_m d_m^2)/F^2. Please rephrase the derivation so that the normalisation by 2E is explained clearly.","section":"§2.1, Eq. (6)"},{"comment":"The values of µ1 and µ2 used in the power-law experiments are not stated in the text or the figure captions of Figs. 8-10; the later mention of µ = 2, 3, 4 should be made explicit together with the network sizes, the number of communities, and the number of realisations, so that the results are reproducible.","section":"§6.2"},{"comment":"The Dual Projection algorithm is recommended first in the conclusions, but its agglomerative clustering step is described only as 'use agglomerative clustering [14] to join the top and bottom communities'. Please specify the linkage criterion and any weighting used, or provide a precise reference to the implementation, since this is the method the paper ultimately recommends.","section":"§4"},{"comment":"There are minor typographical issues, e.g., 'synthethic' in Section 8 and a duplicated 'References' heading; I assume these will be corrected in revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of Physica A and the core mathematical content is sound. The main risk is that the practical heuristic is derived from, and evaluated on, a single synthetic family that coincides with the null model of QP, while the real-data support is only indirect. I would like the revision to either broaden the experiments or explicitly restrict the claims. The H/C definitional error is easy to fix and should not be the basis of a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something genuinely useful: it defines a modularity QP for projected bipartite networks, where the null model rewires the original bipartite graph and then projects. That's the right null model for the problem, and the derivation, including the Louvain gain formula and the proof that projection commutes with taking induced subgraphs, is correct. The synthetic experiments are also decently designed: 100 realizations, 10 Louvain runs, bootstrap confidence intervals, and three metrics.\n\nOn the soft side, the paper overreaches a bit. The headline heuristic — use the bipartite graph when µ2 > µ1, project when µ2 ≤ µ1 — is derived entirely from synthetic graphs generated by a configuration model with planted communities. That generative process is close to QP's own null model, so the benchmark is favorable to QP. Real networks will often have nestedness, degree correlations, or heterogeneous community sizes, and the paper does not test those. Also, on the four real datasets the exponents are read off by eye from Figure 11, with no fitting procedure, so the ordering µ1 vs µ2 can be fragile. Table 1 shows that Standard and Projected produce almost identical partitions on the real data, which tempers the practical importance of QP. And the recommendation to use Dual Projection as the default comes mostly from synthetic results; on real data it produces the most divergent partitions.\n\nThere is also a small but real text error in Section 5: the paper says H is reduced when two nodes from the same target community are assigned to different communities. That statement describes a completeness failure, not a homogeneity failure. It doesn't affect the experiments, but it should be fixed.\n\nNone of this breaks the central contribution. The QP definition is mathematically sound and likely to be useful to practitioners. The heuristic is a reasonable rule of thumb, but should be presented as a starting point rather than a proven law. The main missing piece is code/data for reproducibility.\n\nThis paper deserves a serious referee. I'd send it out, and ask the referee to focus on the domain of validity of the heuristic and the real-data exponent estimation. For a reading group on community detection, it would be a good case study in how to (and how not to) justify a practical rule.","headline":"A mathematically sound new modularity for projected bipartite networks, paired with a useful but under-tested heuristic; worth refereeing.","tokens_in":13368,"tokens_out":4567,"would_cite":true,"duration_ms":47453,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","05C82","68R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new modularity function for projected bipartite networks, using a null model that rewires the original two-mode graph before projecting, recovers planted communities at least as well as standard modularity and supports a simple…","keywords":["bipartite networks","community detection","projected modularity","network projection","power-law degree distributions","null model","two-mode networks","synthetic benchmarks"],"falsifier":"Take a two-mode network with known planted communities, set the projected-away degree distribution to be heavier-tailed than the kept one ($\\mu_2 > \\mu_1$), but arrange the high-degree nodes so that they connect across communities rather than within them. If a projected method recovers the planted partition while bipartite methods fail, the paper's rule that heavy-tail top nodes force bipartite detection would be wrong; looking for such a counterexample directly tests the mechanism.","tokens_in":12387,"feed_emoji":"🕸️","tokens_out":8393,"duration_ms":84365,"temperature":0.7,"pith_summary":"This paper argues that community detection on projected bipartite networks should use a modularity function whose null model respects how the projection was made. The author proposes $Q_P$, defined by rewiring the original two-mode graph and then projecting, so that a node with many partners is not mistaken for an independent hub but is recognized as the source of a clique. On planted synthetic networks, maximizing $Q_P$ recovers the generative communities at least as well as standard modularity, and on power-law networks the best method depends on which node set has the heavier-tailed degree distribution. The paper's practical conclusion is a simple rule: when the projected-away node set has the heavier tail, keep the bipartite graph; otherwise project and use $Q_P$.","feed_headline":"New modularity metric fixes community detection on projected networks","feed_subtitle":"A projection-aware modularity QP tells whether to cluster the bipartite graph or its projection.","key_machinery":"The central object is the projected modularity\n$$Q_P=\\frac{1}{2E}\\sum_{ij}A_{ij}\\,\\delta(c(i),c(j))-\\sum_{ij}\\frac{q_i q_j}{$F^{2}$}\\,\\delta(c(i),c(j)),$$\nwhere $A_{ij}=\\sum_m B_{im}B_{jm}$ is the weighted projection, $q_i$ is the bipartite degree of the kept node, $d_m$ is the degree of the projected-away node, $F=\\sum_i q_i=\\sum_m d_m$, and $2E=\\sum_m d_m^2$. The null model rewires the bipartite graph with degrees fixed and then projects, giving expected projected weight $q_i q_j/F^2$, so high-degree projected-away nodes contribute as clique sources rather than independent links. The paper also proves that forming the induced bipartite graph on a community partition and then projecting commutes with projecting first and then aggregating, which allows a greedy modularity-maximizing algorithm to evaluate local gains for $Q_P$ efficiently.","core_discovery":"The central claim is that standard modularity on a projected two-mode network uses the wrong null model, and that the correct null model—rewire the bipartite network and only then project—yields a projected modularity $Q_P$ that better matches the communities present before projection. The difference shows up sharply around high-degree nodes: a top node with $n$ links induces about $n^2$ projected edges, so cliques appear that are artifacts of projection; the $Q_P$ null model assigns the correct expected weight to those cliques. The paper demonstrates on synthetic networks with planted communities that optimizing $Q_P$ recovers the target partition with accuracy comparable to the best bipartite methods, and that when the degree distributions are power laws the optimal strategy flips according to the exponents: if the projected-away side has a heavier tail ($\\mu_2 > \\mu_1$), projection hides structure and bipartite detection wins; if the kept side has the heavier tail ($\\mu_2 \\le \\mu_1$), projection is safe and $Q_P$ performs best. These patterns are confirmed on four real two-mode networks without ground truth by comparing partitions across algorithms.","pith_inferences":["One natural extension is to replace the visual inspection of degree exponents with a maximum-likelihood estimator, making the $\\mu_2$-versus-$\\mu_1$ rule a fully automated model-selection criterion for real networks.","The $Q_P$ null model should extend to weighted projections with more sophisticated weighting than the simple co-occurrence count; testing it on recommendation-style weighted projections would show whether the clique correction remains the dominant effect.","A sharp testable prediction follows from the paper's mechanism: in a two-mode network with $\\mu_2 > \\mu_1$, the projected graph's clique structure should inflate standard modularity of the planted partition even as recovery fails, so comparing $Q_P$ and standard $Q$ on the planted partition should reveal a characteristic gap.","The commutativity result suggests $Q_P$ could be plugged into hierarchical or overlapping community detection without changing the null model, since induced subgraphs and projection can be interchanged at any resolution."],"forward_implications":["Projection does not destroy community structure for Poisson-degree bipartite networks: communities found on the projection agree with the planted ones about as well as communities found on the full bipartite graph.","When the projected-away node set has a heavier-tailed degree distribution, projection fabricates cliques that hide real communities, so the bipartite graph itself should be clustered.","When the kept node set has the heavier tail, projected methods—$Q_P$ or even standard modularity—recover the planted communities well, so projection is a safe simplification.","A dual-projection approach that clusters both projections and merges them by bipartite modularity gives the closest overall match to the target structure, with higher completeness than other methods.","Even when only the projected graph is available, optimizing standard modularity produces partitions nearly as good as optimizing $Q_P$, so practitioners are not locked into a specialized algorithm."],"supporting_citations":[{"why":"Provides the degree-preserving construction of synthetic bipartite networks and the result that projection preserves the bottom degree distribution when $\\mu_2 > \\mu_1$, which grounds the paper's heuristic.","marker":"[29]"},{"why":"Defines the bipartite modularity $Q_B$ used by the bipartite algorithms and by the dual-projection merging step.","marker":"[23]"},{"why":"Supplies the fast greedy modularity-maximizing procedure that is modified to optimize $Q_P$ and $Q_B$.","marker":"[34]"},{"why":"Supplies the bi-partitioning variant that optimizes bipartite modularity, used as one of the main bipartite algorithms.","marker":"[35]"},{"why":"Introduces the dual-projection strategy of clustering both projections and merging them, which the paper extends by using $Q_P$ in the projection steps.","marker":"[13]"},{"why":"Provides the planted-community-with-mixing benchmark design that the synthetic networks emulate, including the noise parameter $p$.","marker":"[20]"},{"why":"Identifies biclique communities formed by shared partners, the clique-hiding mechanism that motivates the new null model.","marker":"[11]"}],"fun_headline_variants":["New null model fixes modularity for projected bipartite networks","Projection-aware modularity reveals true clusters in two-mode graphs","Why projection creates false cliques: a modularity correction","Rewire before projecting: better modularity for bipartite networks","Degree-tail heuristic chooses best method for bipartite community detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that real bipartite communities look like the synthetic benchmark: $C$ planted components with a fraction $p$ of edges rewired uniformly at random and independent power-law degree sequences whose exponents can be read off the data; if real networks contain degree correlations, nestedness, or heterogeneous community sizes, the $\\mu_2$-versus-$\\mu_1$ rule can break.","fun_headline_variants_meta":{"raw":{"variants":["New null model fixes modularity for projected bipartite networks","Projection-aware modularity reveals true clusters in two-mode graphs","Why projection creates false cliques: a modularity correction","Rewire before projecting: better modularity for bipartite networks","Degree-tail heuristic chooses best method for bipartite community detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1388,"prompt_tokens":881,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":424}},"tokens_in":497,"tokens_out":507,"duration_ms":6661,"temperature":1.0,"reasoning_tokens":424,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:41:21.435529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-mode network with known planted communities, set the projected-away degree distribution to be heavier-tailed than the kept one ($\\mu_2 > \\mu_1$), but arrange the high-degree nodes so that they connect across communities rather than within them. If a projected method recovers the planted partition while bipartite methods fail, the paper's rule that heavy-tail top nodes force bipartite detection would be wrong; looking for such a counterexample directly tests the mechanism.","supporting_citations":[{"cited_title":"Vasques Filho, D","cited_arxiv_id":null,"evidence_quote":"Provides the degree-preserving construction of synthetic bipartite networks and the result that projection preserves the bottom degree distribution when $\\mu_2 > \\mu_1$, which grounds the paper's heuristic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the bipartite modularity $Q_B$ used by the bipartite algorithms and by the dual-projection merging step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fast greedy modularity-maximizing procedure that is modified to optimize $Q_P$ and $Q_B$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the bi-partitioning variant that optimizes bipartite modularity, used as one of the main bipartite algorithms."},{"cited_title":"Melamed, Community structures in bipartite networks: A dual- projection approach, PloS one 9 (5) (2014) e97823","cited_arxiv_id":null,"evidence_quote":"Introduces the dual-projection strategy of clustering both projections and merging them, which the paper extends by using $Q_P$ in the projection steps."},{"cited_title":"Lancichinetti, S","cited_arxiv_id":null,"evidence_quote":"Provides the planted-community-with-mixing benchmark design that the synthetic networks emulate, including the noise parameter $p$."}],"review_version":1}