{"id":"36310dcc-3fc2-4557-99b5-21a5d9a978e4","arxiv_id":"2412.05305","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A clustering framework combining MDL and genetic optimization is proposed, but its own tables contradict the claim of consistent superiority.","lead":"This paper proposes a clustering method that combines minimum description length, genetic algorithms, and ensemble consensus, and reports that it outperforms six classic baselines on thirteen datasets. The reported results do not consistently support the claim, and the method's equations and statistical evidence contain serious gaps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tables 2–5 contradict the central 'consistently outperforms' claim: on balance_scale NMI and several other dataset–metric cells, every reported variant of the proposed method is below k-means or FCM.","rationale":"The paper's advertised contribution is not a derivation or a construction; it is an empirical claim of uniform superiority. The data that would establish that claim are the tables, and those tables contain direct counterexamples. This is the most load-bearing concern because it attacks the conclusion itself rather than a peripheral detail: if the reported values stand, the conclusion is false; if the values are wrong, the paper provides no corrected data, code, or hyperparameters to restore the claim. The reader's weakest_assumption (true-k leakage and k-means-based initialization) identifies a real protocol problem, but my independent concern is even more basic: the paper's own favorable setup already fails to produce consistent superiority. I therefore agree partially with the reader's framing. Secondary issues—malformed equations, undefined variables, missing code—make reproduction impossible but are not needed for the rejection; the internal contradiction is sufficient. I set verdict_should_be to REJECT because the central claim is directly falsified by the reported evidence.","tokens_in":63,"tokens_out":8308,"duration_ms":110710,"concrete_test":"Using only the values printed in Tables 2–5, tabulate every dataset–metric cell in which each proposed variant (ABMDLGAO, EPMDLGAO, EPAFGAO) has a mean below any baseline (k-means, Single/Average/Complete/Ward linkage, FCM). Under the paper's own paired t-test description, a significant loss at p<0.05 on any dataset would falsify 'consistently outperforms'. If raw outputs are available, rerun the 100 trials with the reported settings and compute paired tests. A focused check: re-evaluate balance_scale NMI and Nionosphere accuracy exactly as described; if the printed values reproduce, the central claim is refuted regardless of any reanalysis.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that the Genetic MDL framework 'consistently outperforms traditional clustering methods' on thirteen benchmark datasets. The paper's own experimental tables contradict this. No aggregated 'Genetic MDL' result is defined; the three optimization variants ABMDLGAO, EPMDLGAO, and EPAFGAO are reported as separate methods. Under any reasonable reading of 'consistently', the proposed approach must beat k-means, the four linkage methods, and FCM on most or all datasets. It does not. In Table 2, balance_scale NMI is 12.39 for k-means and 11.63 for FCM, while the three variants score 9.69, 9.28, and 8.26. On Nbalance_scale, k-means NMI is 11.51 while ABMDLGAO is 8.85 and EPMDLGAO 8.47. In Table 4, accuracy on Nionosphere is 70.66 for k-means but 68.82 for ABMDLGAO; on Nwine, k-means is 96.63 while ABMDLGAO is 88.79 and EPAFGAO 67.45. The 'improved stability' claim is also unsupported: Table 2 reports NMI means with standard deviations as large as 6.06 ± 21.91 (Nbreast) and 26.57 ± 40.53 (NYeast). Finally, Tables 6–9 report signed rank scores without stating an aggregate win/loss count or a paired significance test, so no statistical support for superiority is provided. The protocol also leaks the true cluster count into every run ('For all datasets, the number of clusters and the true labels of the samples were known beforehand'), and the agreement matrix plus initial solution are k-means based, weakening the advertised unsupervised setting. But the decisive problem is that even under the authors' own favorable setup the results do not support the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaptiveMDL-GenClust, a clustering framework that combines an ensemble-based initial solution with the Minimum Description Length (MDL) principle, attribute-weighted description length (AWDL), and three genetic-algorithm optimization stages (ABMDLGAO, EPMDLGAO, EPAFGAO). The authors claim that this framework consistently outperforms k-means, four hierarchical linkage methods, and fuzzy c-means on thirteen UCI benchmark datasets, with higher accuracy, improved stability, and reduced bias. The manuscript provides algorithmic outlines, mathematical equations for the proposed objective functions, and experimental tables with means and standard deviations over 100 runs, plus a ranking analysis based on a t-test. The evaluation uses four metrics: accuracy, NMI, Fisher score, and ARI.","tokens_in":19311,"tokens_out":5256,"duration_ms":44419,"significance":"If fully supported, a clustering framework that automatically adapts to dataset structure and outperforms standard baselines across diverse benchmarks would be a meaningful contribution to evolutionary and MDL-based clustering. The use of thirteen datasets and four validation metrics, with 100-run averages, is a reasonable evaluation scale, and the idea of combining MDL-style description length with an ensemble initialization is worth investigating. However, the manuscript as written does not establish the claimed result: the headline claim is directly contradicted by the paper's own tables, the method is not reproducible from the provided equations and missing hyperparameters, the experimental protocol relies on true labels and true cluster counts, and the statistical analysis is incomplete. No code or data are provided. The potential significance is therefore not realized in the current manuscript.","major_comments":[{"comment":"The central claim that the proposed framework 'consistently outperforms' traditional clustering methods is contradicted by the paper's own numbers. In Table 2, balance_scale NMI is 12.39 for k-means and 11.63 for FCM, while ABMDLGAO, EPMDLGAO, and EPAFGAO score 9.69, 9.28, and 8.26, respectively. On Nbalance_scale, k-means NMI is 11.51, above all three proposed variants. Table 4 shows Nwine accuracy of 96.63 for both k-means and FCM, whereas ABMDLGAO is 88.79 and EPAFGAO is 67.45. The 'improved stability' claim is also unsupported: Table 2 reports standard deviations such as 26.57 ± 40.53 for ABMDLGAO on NYeast and 6.06 ± 21.91 for EPAFGAO on Nbreast. The paper never defines an aggregate 'Genetic MDL' result, so no reasonable reading of the tables supports consistent superiority.","section":"Experimental Results, Tables 2-5"},{"comment":"The evaluation protocol states that 'the number of clusters and the true labels of the samples were known beforehand,' and this knowledge is used to set the number of clusters and to compute accuracy, NMI, and ARI. This is a supervised setting, not the unsupervised clustering setting advertised in the abstract. Moreover, the Agreement Matrix (Figure 3) and the initial solution C0 (Figure 4) are built using the same k-means algorithm that later serves as a baseline, so the comparison is not independent of the ensemble construction. The paper provides no experiments for the fully unsupervised case in which k is unknown, despite claiming adaptability and reduced dependency on initial inputs.","section":"Data Sets, Table 1"},{"comment":"Several central equations are malformed or contain undefined symbols, making the method irreproducible. Equation (2) misindexes the normalization formula: X_ij is described both as the jth sample and as the value of the jth attribute of the ith sample. Equation (9) omits the weights w_j on all terms after the first absolute difference. Equation (12) defines probabilities only for the cases max(A_iq) < max(A_i) and max(A_iq) = max(A_i), leaving the case max(A_iq) > max(A_i) undefined, while Eq. (13) defines V_p = 1 − V without defining V. Equation (20) uses inconsistent notation in the NMI definition, and Eq. (22) writes ARI with symbols M, C, and combinatorial terms that are not formally defined. Without corrected equations, the proposed objective functions cannot be implemented or checked.","section":"The Proposed Method, Eqs. (2), (9), (12), (13), (16), (20), (22)"},{"comment":"The genetic algorithm itself is not described. The paper lists three fitness functions (ABMDLGAO, EPMDLGAO, EPAFGAO) but provides no details on population initialization, selection, crossover, mutation, elitism, or termination criteria beyond 'converged or the predefined number of iterations NK is reached.' No values are given for NK, ensemble size, random subset fraction e, or any GA parameter. The 'quasi-code' in Figure 2 only says 'Use the Genetic algorithm to start to optimize,' which is not an algorithm specification. Because the central claim depends on this optimization, this omission is load-bearing.","section":"Genetic algorithm Multiple Optimization framework and Figure 2"},{"comment":"The statistical analysis is incomplete. The text describes a paired t-test, but Tables 6-9 report only integer 'priorities' ranging from -8 to 8, with no test statistic, p-value, or confidence interval, and no aggregate win/loss counts across the 13 datasets. The claimed 'T-test results' therefore do not demonstrate that any difference is statistically significant, and the tables cannot be independently verified from the stated formulas.","section":"Algorithm ranking, Tables 6-9"}],"minor_comments":[{"comment":"The paper uses 'quasi-code' where 'pseudo-code' is standard; please correct this in the text and figure captions.","section":"Throughout"},{"comment":"Figure 1 is mentioned in the text but its diagram is not actually included in the manuscript, so the described framework cannot be visually inspected.","section":"Figure 1"},{"comment":"The dataset name 'Nionosphere' likely refers to 'Ionosphere'; please use the standard UCI dataset name.","section":"Table 1"},{"comment":"There are spelling inconsistencies: 'F-mesture', 'F − meature', and 'Fisher standard' all appear; please unify the terminology.","section":"Equation (18) and surrounding text"},{"comment":"The NMI formula uses undefined symbols such as p_i1, p_i, and p_j; please restate with standard contingency-table notation and define all terms.","section":"Equation (20)"},{"comment":"Reference [40] spells the author as 'Graunwald'; the correct spelling is 'Grünwald'.","section":"References"},{"comment":"The pseudo-code contains a typo, 'Patr1', which should be 'Part1'.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"I concur with the reject recommendation, but my reasoning is grounded in the manuscript itself rather than in the reader's framing. The paper's own tables contradict the abstract's 'consistently outperforms' claim, the method is not reproducible from the equations and missing hyperparameters, and the experimental protocol leaks true labels and cluster counts. These are load-bearing problems that cannot be fixed with a minor revision; a rewrite with corrected equations, full hyperparameter disclosure, code/data release, and an honest comparison in a genuinely unsupervised setting would be required. I also note that the manuscript draws heavily on the authors' prior work on clustering ensembles; this is not itself a problem, but the novelty relative to those works is never stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the central claim is false by the paper's own numbers. In Table 2, k-means beats every proposed variant on balance_scale NMI (12.39 vs 9.69/9.28/8.26), and FCM beats them on Nionosphere (15.83 vs 14.65). Reversals like these appear on several datasets across Tables 2–5, so \"consistently outperforms\" in the abstract and conclusion is not supported.\n\nWhat is genuinely new is the specific assembly: an agreement matrix from an ensemble, an attribute-weighted description length cost, and three genetic-optimization phases. That exact combination is not in the cited references. But the novelty is thin: each component is established, the baselines are old and few, and nothing is shipped. No code, no hyperparameters, no reproducible protocol. The paper also never defines an aggregate \"Genetic MDL\" result; the three variants are reported separately, so \"the method\" is ambiguous.\n\nThe soft spots are serious. First, the evaluation contradicts the headline, which is decisive. Second, the method section is not reproducible: Equations 2, 9, 13, 16, 20, and 22 are malformed or use undefined variables (V in Eq. 13 is never defined; the NMI formula is garbled). Third, the unsupervised framing leaks: the number of clusters is set to the true class count for every dataset, and the agreement matrix and initial solution are built with k-means, the same algorithm used as a baseline. Fourth, Tables 6–9 report signed rank scores without aggregate win/loss counts or a paired significance test, so no statistical conclusion can be drawn.\n\nWho is this for? Someone working on ensemble clustering or evolutionary clustering might skim the framework description, but they cannot use the results. With code and a corrected, honest aggregate comparison, a narrower version of the paper might be worth reviewing. As written, the central empirical claim fails on its own data. I would desk reject it, not send it to reviewers.","headline":"The paper's own tables contradict its headline claim, and without code or working equations the contribution is not assessable; desk-reject rather than send to referees.","tokens_in":19839,"tokens_out":3750,"would_cite":false,"duration_ms":34427,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A genetic algorithm guided by attribute-weighted MDL consistently beats six classical clustering methods on thirteen benchmarks.","keywords":["minimum description length","genetic algorithm","clustering","normalized mutual information","ensemble clustering","attribute weighting","adjusted Rand index"],"falsifier":"Re-run the thirteen-dataset comparison with the cluster count estimated by the MDL criterion itself instead of the known true class count; the central claim stands only if the Genetic MDL advantage over k-means and linkage baselines persists when the number of clusters is not supplied.","tokens_in":18638,"feed_emoji":"🧬","tokens_out":5485,"duration_ms":46753,"temperature":0.7,"pith_summary":"This paper tries to establish that a clustering method can be made markedly more accurate and stable by optimizing not a similarity measure but the compressed description length of the data, and that a genetic algorithm can carry out that optimization effectively. The proposed Genetic MDL framework generates a starting partition from an ensemble of k-means runs, then refines it with three genetic phases guided by an attribute-weighted MDL criterion. On thirteen benchmark datasets, the paper reports that the method beats k-means, four hierarchical linkage variants, and fuzzy c-means on all four validation metrics, averaged over 100 independent runs. If this holds, clustering would gain a relatively tuning-light, data-adaptive recipe that is less hostage to initialization and feature scaling.","feed_headline":"Genetic MDL clustering beats six classics on 13 datasets","feed_subtitle":"MDL-guided genetic search refines an ensemble partition, posting higher accuracy and stability on all four metrics.","key_machinery":"The carrying object is the Attribute Weighted Description Length (AWDL) criterion, written as $L' = \\{S_m, S_d\\}$, where $S_m$ is the weighted sum of cluster-mean magnitudes and $S_d$ is the weighted average absolute deviation of samples from their cluster means, with attribute weights derived from variance. The agreement matrix $A = H H'$ records how often pairs of samples are co-clustered by the input ensemble and supplies both the initial solution and the fitness signal for the genetic phases. ABMDLGAO moves samples between clusters according to agreement-based probabilities, EPMDLGAO does the same with uniform probabilities, and EPAFGAO maximizes a thresholded agreement reward.","core_discovery":"The central claim is that replacing a similarity-only clustering objective with an attribute-weighted description-length objective, optimized by a genetic algorithm, removes much of the bias and initialization dependence of conventional clustering. Concretely, the paper asserts that its Genetic MDL framework, initialized by a consensus partition from repeated k-means runs and refined by three genetic phases named ABMDLGAO, EPMDLGAO, and EPAFGAO, consistently outperforms k-means, single/average/complete/Ward linkage, and fuzzy c-means on thirteen datasets according to accuracy, NMI, Fisher score, and adjusted Rand index, averaged over 100 runs.","pith_inferences":["Beyond the paper: since the agreement matrix is built from k-means runs, the framework's view of consensus inherits k-means' biases; swapping in diverse generators such as density- or model-based clustering is a natural test of whether the MDL refinement, rather than the ensemble source, drives the gains.","Beyond the paper: all four metrics are measured with the true number of clusters supplied; a fair unsupervised comparison would estimate the cluster count from the data, for example by minimizing $L'$ itself, and that is where the reported margin could shrink or vanish.","Beyond the paper: the reported absolute scores on several datasets are low, so 'consistently outperforms' is a relative statement; the method may improve on classical baselines without yet being a universal clusterer on hard data."],"forward_implications":["Clustering objectives can be reformulated as compression problems: minimizing $L'$ should yield partitions that are at once simple and faithful, usable in place of k-means inertia or linkage criteria.","Because the framework starts from an ensemble and evolves it, the method should be more stable across random restarts than k-means, whose results vary with initialization.","Variance-based attribute weighting gives the method a built-in feature-relevance adjustment, which should help on datasets with noisy or irrelevant attributes.","The same pipeline can be retargeted to high-stakes domains named in the paper, such as EEG and ECG signal grouping and spatial transcriptomics, where stable clusters feed downstream diagnosis."],"supporting_citations":[{"why":"Supplies the Minimum Description Length principle and its model-selection foundation, which the AWDL criterion extends.","marker":"[40]"},{"why":"Provides the evidence-accumulation idea behind the agreement matrix $A = H H'$ used to form the ensemble consensus.","marker":"[36]"},{"why":"Frames cluster ensembles as a knowledge-reuse mechanism, motivating the generation of the initial clustering solution.","marker":"[9]"},{"why":"Defines k-means, which is both a baseline algorithm and the mechanism used to build the ensemble and the initial solution.","marker":"[13]"},{"why":"Supports the consensus-combination stage of ensemble clustering that the proposed framework builds on.","marker":"[30]"}],"fun_headline_variants":["MDL + genetic search: clustering that adapts and outperforms","Evolutionary MDL clustering: robust, accurate, low bias on 13 sets","Genetic MDL clustering: data-driven, beats six classic methods","MDL genetic algorithm: clustering with less bias, higher accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison hands every algorithm the true number of classes for each dataset, and the paper says so explicitly in the Data Sets section, so the reported robust and adaptive performance is not tested in the fully unsupervised setting the title advertises.","fun_headline_variants_meta":{"raw":{"variants":["MDL + genetic search: clustering that adapts and outperforms","Evolutionary MDL clustering: robust, accurate, low bias on 13 sets","Genetic MDL clustering: data-driven, beats six classic methods","MDL genetic algorithm: clustering with less bias, higher accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001721,"raw_usage":{"total_tokens":6785,"prompt_tokens":898,"completion_tokens":5887,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":5823}},"tokens_in":514,"tokens_out":5887,"duration_ms":32900,"temperature":1.0,"reasoning_tokens":5823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:45:17.926320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the thirteen-dataset comparison with the cluster count estimated by the MDL criterion itself instead of the known true class count; the central claim stands only if the Genetic MDL advantage over k-means and linkage baselines persists when the number of clusters is not supplied.","supporting_citations":[{"cited_title":"Advances in Minimum Description Length: Theory and Applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the Minimum Description Length principle and its model-selection foundation, which the AWDL criterion extends."},{"cited_title":"Combining Multiple Clusterings using Evidence Accumulation,","cited_arxiv_id":null,"evidence_quote":"Provides the evidence-accumulation idea behind the agreement matrix $A = H H'$ used to form the ensemble consensus."},{"cited_title":"and Ghosh J","cited_arxiv_id":null,"evidence_quote":"Frames cluster ensembles as a knowledge-reuse mechanism, motivating the generation of the initial clustering solution."},{"cited_title":"Some methods for classification and analysis of multivariate observations,","cited_arxiv_id":null,"evidence_quote":"Defines k-means, which is both a baseline algorithm and the mechanism used to build the ensemble and the initial solution."},{"cited_title":"Graph-based consensus clustering for class discovery from gene expression data,","cited_arxiv_id":null,"evidence_quote":"Supports the consensus-combination stage of ensemble clustering that the proposed framework builds on."}],"review_version":1}