{"id":"92e30887-d1a5-41fe-9ce2-046955156a9a","arxiv_id":"1908.11818","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A community detection algorithm that partitions DAG nodes into antichains of similar nodes using a modularity-style 'siblinarity' score.","lead":"This paper introduces a new way to detect communities in directed acyclic graphs, where connections flow one way in time or hierarchy. The method groups nodes that are not directly connected but share the same neighbors, preserving the order of the graph while finding similar items.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Siblinarity null model in Eq. (5) is an unvalidated substitution for the DAG-configuration baseline that the paper itself derives in Eq. (A6); the objective may be biased.","rationale":"The reader's weakest-assumption analysis and my stress-test converge on the same point: the null model in Eq. (5) is the least secure link in the argument. The paper defines siblinarity as observed overlap minus an expected overlap, but the expected value is taken from the configuration model of the derived graph Ã=AAᵀ, not from a random DAG ensemble with the same degree sequence. The authors themselves derive a DAG-specific approximation in Eq. (A6) and explicitly choose not to use it, without giving a substantive reason. This is not an internal contradiction, but it is a real gap: the objective function's behaviour, and therefore the communities it selects, may depend on an arbitrary baseline. If the null were replaced by a more DAG-faithful one and the planted-field recovery in Section 4.2 or the Cora topic homogeneity in Section 4.4 disappeared, the central claim would be seriously weakened. Conversely, if the results were robust to that replacement, the concern would be resolved. The proposed test is direct and computational, using the paper's own data and models, so it would settle the issue. I am not proposing to reject the paper: the algorithm is clearly specified, the antichain constraint is enforced, and the empirical demonstrations are suggestive. But the unvalidated null model is exactly the kind of condition that should be checked before the central claim is accepted. Hence the verdict remains CONDITIONAL, in agreement with the reader.","tokens_in":32406,"tokens_out":20002,"duration_ms":183965,"concrete_test":"Recompute the Section 4.2 Price-model experiment (and the Cora analysis) with λ=1, replacing the null term κ_nκ_m/W in Eq. (5) with the DAG configuration-model expectation from Eq. (A6). Keep the Louvain optimisation and antichain constraint unchanged. If the field-diversity advantage of siblinarity over height/depth partitions is substantially reduced or reversed, the central claim depends on the specific null; if the diversity results are essentially unchanged, the null choice is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that maximizing siblinarity finds antichain communities of similar nodes. Siblinarity in Eq. (5) is S(A)=Σ(Ã_nm − κ_nκ_m/W), where Ã=AAᵀ for successors (or AᵀA for predecessors). The subtraction is supposed to be the expected neighbourhood overlap under a null model. But the paper uses the configuration-model expectation on the derived second-neighbour graph Ã, not on the original DAG. In Appendix A the authors derive a different null, Eq. (A6), for the DAG configuration model preserving in/out degrees: E(|N(n)∩N(m)|)≈((⟨k_in²⟩−1)k_out(n)k_out(m))/(⟨k_in⟩|E|), and then set it aside without justification. These two baselines differ in how they weight high in-degree successors; κ_nκ_m/W can substantially over- or under-penalize pairs whose shared neighbours are popular, changing which antichain partitions are optimal. Since the empirical support for 'finds similar nodes' comes from the same objective, an arbitrary null makes the claim underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes \"siblinarity\", a quality function for partitions of a DAG into antichains. For each community, it sums over node pairs the number of common successors (or predecessors) minus the configuration-model expectation computed on the derived two-step graph (Eq. 5). Maximizing this function with a Louvain-style heuristic is claimed to yield antichain communities whose nodes are both order-incomparable and topically similar. The method is tested on a space-time lattice model, a Price model with planted fields, the Florida Bay food web, the hep-th citation graph, and the Cora citation network. Comparisons with height/depth layering and with modularity communities show that the siblinarity partitions are smaller and more homogeneous in field/topic while respecting the partial order.","tokens_in":32607,"tokens_out":12492,"duration_ms":117117,"significance":"If the method works as described, it is a useful addition to the small literature on DAG-aware clustering: it treats the absence of a directed path as a signal rather than a nuisance, and it transposes bibliometric similarity into a community-detection objective. The paper is commendably concrete: code and data are released in a Figshare repository, and Appendix B fully enumerates all antichain partitions of a small graph, which helps the reader see exactly what the objective rewards. I found no circularity in the use of field labels or publication dates; they enter only in post-hoc validation. The main reservations are (i) the Louvain coarse-graining does not preserve the stated objective, (ii) the null model is chosen by analogy with modularity rather than derived from a DAG ensemble despite the authors' own Eq. (A6), and (iii) the diversity validations are not size-controlled. These issues are addressable, so the contribution is potentially significant, but the present version overstates the support for the central claim.","major_comments":[{"comment":"The second Louvain phase optimizes a different function from S(A) of Eq. (5). When two antichain communities A_a and A_b are merged, the contribution to the original siblinarity is the sum over pairs (n in A_a, m in A_b) of (AA^T)_{nm} minus the corresponding null term. In the coarse-grained graph H of Eq. (C28), however, the two-step entry (HH^T)_{ab} equals the sum over successor communities c of (sum over n in A_a, s in c of A_{ns})(sum over m in A_b, t in c of A_{mt}), which includes products of edges to distinct successors s and t whenever s and t lie in the same community c. This is not equal to the original sum over common successors. Consequently, merges decided at later levels need not increase S(A) on the original DAG, and the final partition need not be a local optimum of the stated objective. Please either prove that the coarse-graining preserves S(A), compute Delta-S for candidate merges directly from the original graph, or benchmark the heuristic against exact optimization on small instances. The worked example in Appendix B checks the exact objective on one small graph but does not test the Louvain phase itself.","section":"Section 2.4 and Appendix C, Eq. (C28)"},{"comment":"The null model used in Eq. (5) is the configuration model on the derived two-step graph ~G, not a null model for the original DAG. In Appendix A the authors derive the DAG-configuration expectation E(|N(n)∩N(m)|) ≈ ((⟨k_in²⟩−1) k_out(n) k_out(m)) / (⟨k_in⟩ |E|) and then state \"we chose not to do so\" without giving a reason. These two baselines are not equivalent: the former subtracts a product of two-step strengths, while the latter weights shared neighbours by their in-degree. Because the optimal antichain partition depends on which baseline is subtracted, the empirical claims about \"similar nodes\" are conditional on an under-justified modeling choice. The paper should either justify the configuration-model-on-~G choice or show that the main findings are robust to using the DAG null of Eq. (A6), and ideally to a resolution-parameter sweep that keeps the mean community size fixed.","section":"Section 2.3, Eq. (5), and Appendix A, Eq. (A6)"},{"comment":"The diversity comparison is confounded by community size. Height and depth antichains are much larger than siblinarity communities, and Shannon diversity of a label distribution tends to increase as more nodes are drawn from a heterogeneous mixture. To support the claim that siblinarity finds same-field or same-topic nodes, the paper needs a size-controlled baseline, for example random antichain partitions with the same community-size distribution, or a plot of diversity as a function of |A|. Fig. 8's restriction to antichains of at least five nodes is a step in the right direction, but Figs. 4-5 do not include such a control and no size-matched baseline is provided anywhere. If the size-matched comparison still shows lower diversity for siblinarity communities, that would substantially strengthen the paper.","section":"Section 4.2, Figs. 4-5, and Section 4.4, Fig. 8"}],"minor_comments":[{"comment":"The term \"weakly connected\" is used to mean \"there is a directed path in either direction\", whereas standard weak connectivity ignores edge direction; consider using \"comparable\" or \"connected by a directed path\" instead.","section":"Section 2.1 and Appendix A"},{"comment":"The self-loop entries of ~A are excluded from the pair sum but are included in κ_n and W; the effect of this choice on the null model should be discussed more explicitly than the brief dismissal of the non-backtracking alternative.","section":"Eq. (5) and Eq. (A7)"},{"comment":"The Fig. 4 caption calls the parameter \"intraconnectivity probability p\" while the text defines φ for the same quantity; the notation should be unified.","section":"Section 4.2, Fig. 4"},{"comment":"For the Cora analysis, report the number and size distribution of siblinarity antichains, including singleton counts, so the reader can assess the size confound directly.","section":"Section 4.4"},{"comment":"Showing that the S(A) value is stable across runs does not demonstrate that the partitions are stable; consider reporting an adjusted Rand index or normalized mutual information between the partitions found in different runs.","section":"Appendix H"},{"comment":"Several axis labels in Fig. 6 appear as garbled sequences such as \"/uni00000013\"; the figure should be regenerated with readable labels.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The three major issues are all addressable within the scope of the manuscript: the Louvain coarse-graining mismatch can be tested on small graphs, the null-model robustness can be checked by recomputing the Price/Cora experiments with Eq. (A6), and the diversity claims can be re-examined with size-matched baselines. If the qualitative conclusions survive those checks, the paper would make a solid contribution to DAG-specific community detection. I do not see a circularity problem, since the field labels and dates are used only for post-hoc validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What’s actually new: siblinarity—maximizing a modularity-like objective on a second-neighbor graph while constraining all communities to be antichains. That is a real departure from both ordinary community detection and graph layering, and it matches the intuition that in DAGs similar nodes are often precisely those that cannot be directly linked. The paper does that well. The synthetic Price model with planted fields gives clean results: siblinarity antichains are much less diverse than height/depth partitions, and the food web and citation examples look sensible. They also ship code and data, and the comparisons against standard methods are fair.\n\nSoft spots, in proportion. The null model is the real question. Equation (5) uses the configuration-model expectation on the derived second-neighbor graph, κnκm/W, but in Appendix A the authors themselves derive a DAG configuration-model baseline (A6) preserving in/out degrees, then set it aside without a substantive justification. That is an ad hoc choice, and the stress-test worry is legitimate: the two baselines weight popular shared neighbors differently, so the objective could in principle favor partitions that are artifacts of the null rather than of true similarity. However, I don’t think it’s fatal. The planted-field experiments provide a direct check: the method recovers field structure despite the null choice, which suggests the bias is not overwhelming. Still, a referee should ask for either a justification of κnκm/W or a comparison against the DAG-derived null on at least the synthetic models. Minor points: the diversity differences are presented without significance testing, and the Louvain coarse-graining can introduce cycles, so the antichain guarantee on the derived graph does not strictly carry over—worth a sentence or two but not a blocker.\n\nCitation pattern looks honest. They discuss prior DAG community detection (Speidel et al.) and layering methods, and they don’t oversell novelty. The modularity analogy is acknowledged.\n\nWho this is for: network scientists working on citation networks, food webs, package dependency graphs, or any DAG where edges encode order and similarity is carried by shared neighborhoods. It would be a useful reading-group paper too.\n\nRecommendation: send it to peer review. The null-model concern is real but addressable, and the core idea is solid and sufficiently different from existing methods. A serious referee could ask for targeted experiments rather than a rewrite. I would engage with it.","headline":"A genuinely new and useful DAG community detection method that respects partial orders; the null-model choice is the main weak spot, but the empirical results carry the case.","tokens_in":33150,"tokens_out":1950,"would_cite":true,"duration_ms":21159,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper defines siblinarity, a modularity-like score that partitions a directed acyclic graph into antichain communities of nodes with unusually high neighbour overlap.","keywords":["community detection","directed acyclic graphs","antichains","siblinarity","neighbourhood overlap","bibliographic coupling","co-citation","DAG layering"],"falsifier":"Generate a planted-field citation DAG and shuffle field labels among nodes with identical degree sequences, then run the siblinarity optimisation; if the communities still show near-minimal topic diversity, the apparent topical coherence is an artefact of the null model rather than evidence that the method finds similar nodes.","tokens_in":32195,"feed_emoji":"🕸️","tokens_out":11349,"duration_ms":98384,"temperature":0.7,"pith_summary":"Directed acyclic graphs (DAGs) carry an intrinsic order: two papers that reach similar results at the same time cannot cite each other, and two software packages with the same role share dependencies rather than link to each other. The paper argues that meaningful communities in such networks should be antichains—sets of nodes with no directed path between any two members—because that respects the order built into the data. To keep these antichains useful, it introduces siblinarity, a modularity-like score that rewards a community when its members share more neighbours than a random baseline with the same degrees would predict. Maximising siblinarity under the antichain constraint produces communities that are both order-respecting and topically similar, as shown on planted-field citation models, real citation networks, and a food web. The result is a way to cluster DAG nodes that is neither standard layering nor standard community detection but combines the virtues of both.","feed_headline":"Siblinarity partitions DAGs into antichains of similar nodes","feed_subtitle":"For citation networks and food webs, it groups comparable papers, packages, and competing species by shared neighbours.","key_machinery":"An antichain partition and the siblinarity objective $$S(A)=\\sum_{A\\in\\mathcal A}\\sum_{n\\in A}\\sum_{m\\in A\\setminus n}\\left(\\tilde A_{nm}-\\frac{\\kappa_n\\kappa_m}{W}\\right),$$ with $\\kappa_n=\\sum_m \\tilde A_{nm}$ and $W=\\sum_{n,m}\\tilde A_{nm}$. The matrix $\\tilde A$ counts two-step walks on the DAG—one step forward and one back for successor overlap, one back and one forward for predecessor overlap—so its off-diagonal entries are common-neighbour counts. The subtracted term is the configuration-model expectation of that overlap on the derived graph, and the antichain constraint forbids any directed path between nodes in the same community. Maximising $S(A)$ under that constraint is what lets the algorithm turn 'no path between members' from a prohibition into the organising principle; a resolution parameter $\\lambda$ scales the null term, interpolating between small communities of very similar nodes and layer-like partitions.","core_discovery":"The central claim is that a good community in a DAG is an antichain of similar nodes, and that such communities can be found by optimising a modularity-like quantity rather than by layering or by ordinary community detection. The quantity, siblinarity, compares, for every pair of nodes in a community, the observed overlap of their neighbourhoods with the overlap expected from a configuration model built on the derived 'second-neighbour' graph whose adjacency matrix is $\\tilde A = A A^{\\mathrm T}$ for successor overlap or $A^{\\mathrm T} A$ for predecessor overlap. Because the partition is required to be an antichain, no two nodes in a community are connected by a directed path, so the DAG's partial order is respected by construction, while the similarity term ensures the community is not merely a layer but a set of genuinely comparable nodes. On planted-field citation models the siblinarity communities have near-minimal topic diversity, on real citation data they group papers of similar age and topic, and on a food web they group competing species that share prey.","pith_inferences":["A natural testable extension is to apply siblinarity to DAGs where node similarity is also known from independent metadata not used in the construction, such as author teams or software functionality; high agreement would support the claim that common-neighbour overlap captures semantic similarity rather than degree effects alone.","Because the antichain constraint is defined only by absence of a path, the same objective can be used on directed graphs with few cycles, suggesting a route to fixing 'bad' backward citations by merging the two antichain communities that a bad link separates.","If the null model is replaced by an ensemble of random DAGs with the same degree sequence, the siblinarity of a partition could be turned into a statistical z-score, giving a principled way to choose the resolution parameter $\\lambda$ without scanning it manually."],"forward_implications":["In citation networks, siblinarity communities group papers published close in time and within the same topic, so they offer a fairer baseline for comparing citation impact than communities that span many publication dates.","In food webs, the method clusters species at the same trophic level that compete for the same prey rather than mixing predators with their prey, giving an ecological interpretation distinct from standard modularity.","Because the partition must be an antichain, siblinarity communities are guaranteed to respect the partial order of a DAG, a property that ordinary community detection, even on the undirected version of the graph, does not provide.","The resolution parameter allows the same algorithm to reproduce, at one extreme, small communities of highly similar nodes and, at the other, layer-like partitions close to traditional height or depth layering.","On a citation network with known topic labels, siblinarity communities show much lower topic diversity than height-based layers, confirming that the neighbourhood-overlap term, not the antichain constraint alone, is what produces topical coherence."],"supporting_citations":[{"why":"Supplies bibliographic coupling, the shared-bibliography similarity that motivates the successor-neighbour overlap term.","marker":"[8]"},{"why":"Supplies co-citation similarity, the shared-citing-paper measure used for the predecessor-neighbour variant.","marker":"[10]"},{"why":"Provides the matrix formulation that turns directed neighbourhood overlap into the $\\tilde A$ adjacency matrix used in the siblinarity objective.","marker":"[12]"},{"why":"Supplies the configuration-model null model and the path and partial-order definitions underlying the antichain constraint.","marker":"[25]"},{"why":"Supplies the modularity function whose form siblinarity emulates, comparing observed similarity with a null-model expectation.","marker":"[29]"},{"why":"Supplies the fast greedy multilevel optimisation scheme that the paper adapts to maximise siblinarity under the antichain constraint.","marker":"[31]"},{"why":"Supplies the cumulative-advantage citation model that the planted-field test network extends with topic fields.","marker":"[32]"},{"why":"Supplies the cycle-breaking method used to convert real directed networks with a few backward links into DAGs for testing.","marker":"[35]"}],"fun_headline_variants":["Antichain communities for DAGs: respect order, group similar nodes","Siblinarity: modularity for order-respecting DAG communities","Grouping similar nodes in citation networks via siblinarity","DAG communities as antichains of comparable nodes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the configuration-model formula $\\kappa_n\\kappa_m/W$, applied to the derived second-neighbour graph, gives the correct expected neighbourhood overlap for nodes in a DAG, so that any excess overlap above it counts as genuine similarity.","fun_headline_variants_meta":{"raw":{"variants":["Antichain communities for DAGs: respect order, group similar nodes","Siblinarity: modularity for order-respecting DAG communities","Grouping similar nodes in citation networks via siblinarity","DAG communities as antichains of comparable nodes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1653,"prompt_tokens":879,"completion_tokens":774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":700}},"tokens_in":495,"tokens_out":774,"duration_ms":7212,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:05:43.554216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a planted-field citation DAG and shuffle field labels among nodes with identical degree sequences, then run the siblinarity optimisation; if the communities still show near-minimal topic diversity, the apparent topical coherence is an artefact of the null model rather than evidence that the method finds similar nodes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the modularity function whose form siblinarity emulates, comparing observed similarity with a null-model expectation."},{"cited_title":"American documentation 14(1), 10–25 (1963)","cited_arxiv_id":null,"evidence_quote":"Supplies bibliographic coupling, the shared-bibliography similarity that motivates the successor-neighbour overlap term."},{"cited_title":"Journal of the American Society for Information Science 24(4), 265–269 (1973)","cited_arxiv_id":null,"evidence_quote":"Supplies co-citation similarity, the shared-citing-paper measure used for the predecessor-neighbour variant."},{"cited_title":"In: Proceedings of the 14th International Conference on Extending Database Technology - EDBT/ICDT ’11","cited_arxiv_id":null,"evidence_quote":"Provides the matrix formulation that turns directed neighbourhood overlap into the $\\tilde A$ adjacency matrix used in the siblinarity objective."},{"cited_title":"Oxford University Press, Oxford (2010)","cited_arxiv_id":null,"evidence_quote":"Supplies the configuration-model null model and the path and partial-order definitions underlying the antichain constraint."},{"cited_title":"Journal of Statistical Mechanics: Theory and Experiment 2008(10), 10008 (2008)","cited_arxiv_id":null,"evidence_quote":"Supplies the fast greedy multilevel optimisation scheme that the paper adapts to maximise siblinarity under the antichain constraint."},{"cited_title":"J.Amer.Soc.Inform.Sci","cited_arxiv_id":null,"evidence_quote":"Supplies the cumulative-advantage citation model that the planted-field test network extends with topic fields."},{"cited_title":"In: Proceedings of the 2017 ACM on Web Science Conference - WebSci ’17","cited_arxiv_id":null,"evidence_quote":"Supplies the cycle-breaking method used to convert real directed networks with a few backward links into DAGs for testing."}],"review_version":1}