{"id":"f2609016-0dd4-4724-a227-1e42828ae605","arxiv_id":"2508.04034","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"HCE selects the most informative level in a hierarchical clustering by maximizing the entropy of community sizes against the number of communities, and reports strong matches to ground truth on benchmarks.","lead":"This paper proposes a new rule, called HCE, for choosing which level of a hierarchical network clustering to trust: pick the level whose community sizes are most unevenly distributed, balanced against how many communities there are. If it works on real data, it would give researchers a general, domain-agnostic way to find meaningful scales in social and biological networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HCE's size-only objective may select spurious levels on null or imbalanced networks; benchmark matches could reflect dendrogram biases.","rationale":"The reader's verdict of UNVERDICTED is appropriate because the full text is unreadable in this pipeline, leaving the method's derivation and benchmark numbers unverifiable. My stress-test pass identifies a more specific, load-bearing concern than the reader's general worry: HCE's objective is purely a function of community sizes and community count, not of the network's structure. This means the criterion could, in principle, select arbitrary levels on any dendrogram, including ones from pure noise. The reader's weakest_assumption already pointed at the lack of independent justification for the entropy criterion; I agree and sharpen it by noting that the objective discards all edge-level information except what is baked into the dendrogram. The proposed test—running HCE on a null model and an imbalanced planted partition—would settle whether the criterion is biased. Because we still lack access to the manuscript's methods and results, the verdict should remain UNVERDICTED, not REJECT or ACCEPT; the test would determine which direction it should move. I give credit for the method being model-agnostic and scalable in principle, but these advantages do not substitute for a demonstrated link between size-distribution entropy and meaningful scale.","tokens_in":5183,"tokens_out":5331,"duration_ms":69350,"concrete_test":"Run HCE on two controlled cases using the paper's own clustering pipeline: (i) a null network with no planted community structure (e.g., Erdős–Rényi G(n=500, p=0.01) and a geometric random graph), and (ii) a stochastic block model with an imbalanced planted partition (e.g., one community of 80% and four of 5%). If HCE returns k>1 on the null networks, or its selected cut on the SBM does not match the planted partition (measured by NMI), the criterion is not specifically indicative of true structure. Report selected k and NMI for each case; a positive/null mismatch would directly falsify the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that maximizing a trade-off between community-size entropy and community count identifies informative scales. But the HCE criterion operates only on the multiset of community sizes at each dendrogram cut, ignoring edge weights, merge heights, and node labels. Thus, it cannot distinguish a partition that aligns with true modules from any other partition with the same size distribution. No independent argument is given for why a particular entropy/count optimum should correspond to true structure. If the objective favors balanced sizes (as high Shannon entropy does), it will systematically select balanced cuts even when planted communities are imbalanced; if the penalty term instead favors imbalance, it may select spurious uneven partitions in noise. The abstract mentions benchmarks with varying size imbalance, raising circularity risk: if benchmarks reward the same size patterns the objective targets, agreement is baked in. Without a null-model baseline or a first-principles derivation, HCE's selections on real networks are unfalsified, and the reported community alignment could be an artifact of the clustering algorithms' dendrogram shapes rather than evidence for HCE's validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hierarchical Clustering Entropy (HCE), a criterion for choosing informative resolution levels from any hierarchical clustering dendrogram. HCE scores each cut by a trade-off between the entropy of the community-size distribution and the number of communities, and the authors claim it is model-agnostic, scalable, and selects partitions closely aligned with ground truth on LFR and multiscale synthetic benchmarks, while yielding interpretable hierarchies on real social and neuroscience networks.","tokens_in":5363,"tokens_out":3734,"duration_ms":48136,"significance":"If fully substantiated, HCE would be a useful and computationally light addition to the hierarchical community-detection toolbox, since it operates directly on dendrograms and does not require edge-level statistics. The core idea is appealing and falsifiable, and the paper's emphasis on a size-distribution criterion is novel relative to modularity- or likelihood-based approaches. However, the submission as available does not establish the central premise or empirical alignment: no quantitative metrics, baselines, or statistical tests are reported, and the abstract alone cannot support the strength of the claim.","major_comments":[{"comment":"The central claim of 'consistently identifies partitions closely aligned with ground truth' is unsupported by any reported quantity. The abstract gives no NMI/ARI values, no error bars, no baseline comparisons, and no significance testing. This makes the primary empirical assertion unfalsifiable as written and prevents the reader from judging the method's practical value. At minimum, the paper must report standard agreement metrics and compare against a simple size-based baseline.","section":"Abstract, empirical evaluation"},{"comment":"The load-bearing premise is that the cut maximizing a trade-off between community-size entropy and community count marks the informative scale. The abstract asserts this without an independent argument. Because the criterion ignores edge weights, merge heights, and node labels, it cannot distinguish a partition aligned with true modules from any other partition with the same size multiset. A derivation, consistency argument, or a null-model test showing why this specific objective identifies signal rather than artifact is needed.","section":"Abstract, criterion rationale"},{"comment":"The abstract states that the benchmarks vary 'size imbalance' and include both symmetric and asymmetric multiscale models. If the planted ground-truth partitions are themselves size-heterogeneous, then a criterion that rewards size heterogeneity may succeed in part by construction. The paper must demonstrate that HCE does not also select spurious levels on balanced ground-truth communities or on null models with no community structure, e.g., by reporting performance as a function of planted size imbalance.","section":"Synthetic benchmark design"}],"minor_comments":[{"comment":"The phrase 'scales of high structural heterogeneity' is used without a precise definition. This should be formalized or replaced with a quantitative statement.","section":"Abstract, terminology"},{"comment":"The trade-off between entropy and community count is not specified as fixed, tuned, or free. State whether a single default weight is used across all experiments and how sensitive the results are to that choice.","section":"Abstract, parameters"},{"comment":"The supplied manuscript text is heavily corrupted, with most characters replaced by placeholders. If this reflects the submitted PDF, a clean and readable version is essential; equations, figures, and benchmark tables cannot currently be verified.","section":"Full text accessibility"},{"comment":"The abstract does not cite prior work on entropy-based clustering or the LFR benchmark. The body should clearly state the incremental contribution relative to existing scale-selection methods such as modularity-based or persistence-based approaches.","section":"References and positioning"}],"recommendation":"major_revision","confidential_remarks":"The provided manuscript text is largely unreadable due to encoding issues, so I could only evaluate the abstract and the reader's report. The editor should ask the authors for a clean PDF. The circularity concern about the benchmark design is substantive: agreement on size-imbalanced planted communities may be largely a consequence of the HCE objective itself. The revision should include balanced-community controls, null-model baselines, and quantitative metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I could only read the abstract; the full-text rendering in our pipeline is garbage, so everything below the abstract is an inference. That alone forces a provisional verdict.\n\nWhat the paper does, as stated, is neat: it defines Hierarchical Clustering Entropy (HCE), a criterion that scores each cut of a dendrogram using only the multiset of community sizes. The level you pick maximizes a trade-off between size entropy and community count. That is genuinely different from the usual edge-statistic or modularity-based level selection, and if it works, it is a fast, model-agnostic tool for anyone using hierarchical clustering on networks. The abstract is well written and the framing is honest about the method's generality.\n\nThe soft spot is the one the stress-test flags, and I think it is real. The criterion ignores everything except the sizes of the communities at a given cut: no edge weights, no merge heights, no node attributes, no internal structure. So HCE is blind to whether a partition aligns with true modules; it only sees how uneven the sizes are. A high entropy of sizes could come from a spurious cut in noise, and a balanced, meaningful subdivision could be penalized simply for being balanced. The abstract gives no first-principles argument for why maximal size heterogeneity should track ground truth. Then there is the benchmark circularity: the synthetic models deliberately vary size imbalance, which is exactly what HCE rewards. If the planted communities have size distributions that look like the ones HCE likes, agreement might be baked in. The abstract reports no NMI/ARI numbers, no baselines, no error bars, so I cannot tell whether the agreement is impressive or trivial.\n\nNow, fair credit: the idea is not crazy. Entropy-based selection has precedents, but working directly on dendrogram size distributions without edge information is a real angle, and the claimed scalability and domain independence are nice if substantiated. The paper is not incoherent; the concern is under-justification, not nonsense. But with only the abstract in hand, the central claim is unverified and the load-bearing assumption is unexamined.\n\nWho is this for? People building practical community-detection pipelines who want an automatic, parameter-light way to pick a level. That is a useful audience. Whether the paper deserves a referee depends on the full text: if it includes the derivation, the benchmark protocol, and at least one null model showing HCE rejects noise, then yes, it is a legitimate submission. Based on the abstract alone, I would not desk-reject it; the idea is specific enough to warrant a careful reviewer. I just would not cite it or put it in our reading group until I can see the actual numbers.","headline":"Abstract promises a clean, size-only dendrogram-level criterion for choosing informative scales; the idea is worth a look, but the full text is unreadable in this pipeline and the reported evidence boils down to an unquantified claim of ground-truth agreement, with real circularity risk.","tokens_in":5839,"tokens_out":1729,"would_cite":false,"duration_ms":24879,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dendrogram-only entropy criterion selects informative community scales without edge statistics.","keywords":["hierarchical community detection","maximum entropy","dendrogram","renormalization group","multiscale networks","LFR benchmark","modularity","network science"],"falsifier":"Construct an LFR-style network with a planted two-level hierarchy, record HCE's selected level, then rewire edges so the planted top-level communities are equal-sized while keeping connectivity otherwise unchanged; if HCE stops selecting the planted top level, the criterion is tracking size heterogeneity rather than the underlying structure.","tokens_in":5026,"feed_emoji":"🕸️","tokens_out":4639,"duration_ms":64271,"temperature":0.7,"pith_summary":"The paper introduces Hierarchical Clustering Entropy (HCE), a model-agnostic rule for choosing which level of a hierarchical community-detection tree actually carries meaning. Instead of looking at edges, HCE scores each level by a trade-off between the entropy of the community-size distribution and the number of communities, and selects the level where that trade-off is maximized. The paper claims this identifies scales of high structural heterogeneity in any dendrogram, regardless of how the dendrogram was produced. A sympathetic reader should care because this turns scale selection in community detection into a single, parameter-free criterion that could apply broadly across networks and other hierarchical data.","feed_headline":"Entropy rule picks meaningful scales from any community dendrogram","feed_subtitle":"HCE balances community-size entropy against community count and matches planted structure on synthetic benchmarks.","key_machinery":"Hierarchical Clustering Entropy (HCE) is a dendrogram-level score: at each level of a tree of nested partitions, it combines the entropy of the community-size distribution with the number of communities, and the selected scale is the level maximizing that combined objective. The renormalization-group framing in the title ties the chosen levels to successive coarse-grainings of the network, but the operative mechanism is the HCE score applied to whatever dendrogram the user already has.","core_discovery":"Hierarchical Clustering Entropy is a criterion that works directly on the dendrogram, not on edge-level statistics. For each candidate partition level, it forms a score from two ingredients: how spread out the community sizes are, measured by entropy, and how many communities exist. The informative resolution is claimed to be the level where these two terms balance in a way that corresponds to high structural heterogeneity. The paper demonstrates on synthetic benchmarks, including LFR and symmetric and asymmetric multiscale models, that HCE's chosen partitions closely match planted ground truth across varying hierarchy, size imbalance, and noise. On real social and neuroscience networks, the","pith_inferences":["Because HCE consumes only a dendrogram, it should transfer to non-network hierarchical clusterings, such as gene-expression or document trees, where edge statistics are meaningless; the paper does not demonstrate this transfer.","If community-size heterogeneity is the signal, HCE may systematically discount scales with balanced but genuinely distinct communities; a targeted benchmark with planted equal-size communities would test whether the criterion tracks structure or merely size skew.","The entropy-versus-count objective resembles an information-theoretic model-selection score, so one could build a null distribution of HCE values from randomized dendrograms and attach significance to selected levels; the paper does not provide such a statistical test."],"forward_implications":["Any existing hierarchical clustering method can be paired with HCE without modifying the underlying algorithm; the criterion only reads the dendrogram.","Because HCE avoids edge statistics, it applies equally to modularity-based and correlation-based dendrograms, and to any distance metric that produces a hierarchy.","HCE returns multiple informative levels, not just one partition, so it yields a hierarchy of meaningful scales rather than a single cut.","On LFR and multiscale benchmarks, the selected levels align with ground truth under varying hierarchy depth, community-size imbalance, and noise.","Applied to social and brain networks, the selected levels give interpretable modular hierarchies matching known structural and functional organization."],"supporting_citations":[],"fun_headline_variants":["Entropy picks the right scale in any community tree","Dendrogram entropy finds meaningful community levels","Balance entropy and count to read any community dendrogram","Entropy rule finds community scales without edge statistics","Entropy chooses the best community resolution from dendrograms"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The argument stands or falls on the premise that the most heterogeneous distribution of community sizes, rather than any edge-level statistic, is what makes a resolution level informative.","fun_headline_variants_meta":{"raw":{"variants":["Entropy picks the right scale in any community tree","Dendrogram entropy finds meaningful community levels","Balance entropy and count to read any community dendrogram","Entropy rule finds community scales without edge statistics","Entropy chooses the best community resolution from dendrograms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000598,"raw_usage":{"total_tokens":2616,"prompt_tokens":708,"completion_tokens":1908,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1834}},"tokens_in":452,"tokens_out":1908,"duration_ms":17102,"temperature":1.0,"reasoning_tokens":1834,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:54:54.610328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct an LFR-style network with a planted two-level hierarchy, record HCE's selected level, then rewire edges so the planted top-level communities are equal-sized while keeping connectivity otherwise unchanged; if HCE stops selecting the planted top level, the criterion is tracking size heterogeneity rather than the underlying structure.","supporting_citations":[],"review_version":1}