{"id":"15f93428-292d-415a-9db8-a7f76971f713","arxiv_id":"2608.00346","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ensemble variants of deep clustering methods (embedding-averaging and majority voting) achieve the best average ACC, NMI, and ARI on 16 imbalanced binary tabular datasets.","lead":"This paper benchmarks six deep clustering methods on imbalanced tabular datasets and proposes two ensemble strategies that combine cluster assignments across embeddings or across methods. The ensembles rank best on average, but the paper does not compare with supervised classifiers despite claiming deep clustering is a strong alternative.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that deep clustering is a strong alternative to supervised classification is untested: no supervised baseline appears anywhere in the experiments.","rationale":"The paper's central claim, as stated in the abstract, is that deep clustering can serve as a strong alternative to supervised classification under data imbalance. This is a comparative claim and the most load-bearing assertion of the paper. The experimental evidence, however, only benchmarks clustering methods against each other; no supervised classifier is included. The reader's verdict of CONDITIONAL already asks for supervised baselines, but the reader's stated weakest assumption is the KGT post-hoc selection rather than the missing supervised comparison. I agree the KGT selection is a genuine methodological flaw, but the supervised-baseline omission is more fundamental because it directly determines whether the paper's primary significance statement is supported. The proposed concrete test is a direct way to settle this: compare ensembles against simple supervised baselines on the same data. If the ensembles do not compete, the conclusion should be withdrawn; if they do, the claim is validated. The paper does have independent value as a systematic comparison of deep clustering methods on imbalanced tabular data, and the G-CEALS-EE ensemble idea is plausible, so a conditional accept requiring this test remains appropriate. I therefore see no need to change the reader's verdict, hence UNCHANGED, while noting the concern differs from the reader's weakest assumption.","tokens_in":9486,"tokens_out":3581,"duration_ms":35366,"concrete_test":"Run standard supervised classifiers (e.g., logistic regression with class weighting, XGBoost, and a random forest) on the same 16 OpenML datasets and the five controlled class-ratio configurations (1:1, 1:5, 1:7, 1:10, 1:15). Use the same ground-truth labels and evaluate ACC, NMI, and ARI exactly as done for the clustering methods. Compute average ranks across datasets and compare with KGT and G-CEALS-EE. If the supervised classifiers' average ranks are significantly better, the 'strong alternative' claim fails; if the ensembles are competitive or better, the claim gains support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion both assert that 'deep clustering can serve as a strong alternative to supervised classification' in imbalanced settings. This is a central comparative claim, yet the experimental design includes no supervised classifier at all. All eight methods compared (K-means, QGRL, IDC, DCSS, TabClustPFN, G-CEALS, G-CEALS-EE, KGT) are unsupervised; the evaluation tables (Table 2, Figure 1) show only clustering methods against each other. Without measuring the accuracy, NMI, or ARI of standard supervised baselines (e.g., logistic regression with class weights, XGBoost) on the same 16 datasets and the same induced imbalance ratios, the paper cannot support its headline implication that deep clustering is a viable label-free substitute. Even if the ensemble methods outperform other clustering methods, they may still be far less accurate than a supervised classifier; the phrase 'strong alternative' requires direct comparison. The reader's identified concern about KGT's post-hoc member selection on the evaluation set is also valid and could inflate the ensemble's apparent advantage, but it is secondary: fixing that selection issue still leaves the supervised-comparison claim unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates eight clustering methods, including two proposed ensemble approaches, on 16 binary OpenML tabular datasets under natural and artificially induced class imbalance. G-CEALS-EE aggregates G-CEALS cluster assignments across embedding dimensions; KGT applies majority voting to the top three ranked methods (K-means, G-CEALS-EE, TabClustPFN). Performance is measured by ACC, NMI, and ARI. The authors report that the ensemble methods achieve the best average ranks and conclude that deep clustering can serve as a strong alternative to supervised classification in imbalanced settings.","tokens_in":9782,"tokens_out":1796,"duration_ms":18324,"significance":"If the empirical claims hold, the paper provides a useful comparative benchmark of deep clustering methods on tabular data under imbalance, and the two ensemble strategies are simple, generalizable ideas. The study is also honest in reporting mixed per-method strengths (e.g., TabClustPFN's high NMI but poor ACC). However, the headline claim about being a 'strong alternative to supervised classification' is not supported by the experimental design, and the KGT ensemble's composition is selected post hoc on the evaluation data. The paper would be strengthened by adding supervised baselines, a proper validation-based selection protocol, and statistical significance testing.","major_comments":[{"comment":"The abstract and conclusion state that deep clustering 'can serve as a strong alternative to supervised classification' in imbalanced settings, but no supervised classifier is evaluated anywhere in the paper. All eight compared methods are unsupervised. To support this comparative claim, the experiments must include standard supervised baselines (e.g., logistic regression with class weights, XGBoost) on the same 16 datasets and the same imbalance ratios, reporting ACC, NMI, and ARI (or at least ACC) against the same ground-truth labels. Without such a comparison, the paper only supports the weaker claim that ensembles outperform individual clustering methods.","section":"Abstract and Section 5/6"},{"comment":"KGT's membership is selected by ranking the individual clustering methods on the same 16 datasets used for evaluation ('we rank the clustering methods by overall performance and then obtain an ensemble of inferences from the top three methods'). Since K-means, G-CEALS-EE, and TabClustPFN are chosen after seeing their performance on these datasets, the reported KGT advantage is at least partly self-confirming and may not generalize. The authors should select ensemble members on a validation split (or prespecify them based on published benchmarks) and then evaluate on held-out test datasets, or otherwise demonstrate that KGT's superiority is robust to member selection.","section":"§4.1, §5.2, Table 2"},{"comment":"The reported scores are min–max normalized per dataset across methods before averaging, and the paper provides no statistical significance tests. The parentheses in Table 2 appear to be standard deviations, but no pairwise comparisons (e.g., Friedman test with Nemenyi post-hoc, or Wilcoxon signed-rank with multiple-testing correction) are reported. The claim that 'ensemble methods achieve the best performance' is based solely on average ranks. Adding significance tests and reporting raw (unnormalized) average scores would make the comparison more robust and interpretable.","section":"§5.2 and Figure 1"},{"comment":"The definition of ACC in Eq. (4) is written as a maximum over a function n(·) but the notation is incomplete: it suggests the Hungarian optimal mapping, yet the equation as typeset is not a well-defined optimization. The text explains the Hungarian method, but the formal notation should be corrected to make the optimization explicit. This is a minor technical issue but affects a core evaluation metric.","section":"§4.4"}],"minor_comments":[{"comment":"The 50% label-match threshold in Eq. (1) is arbitrary and not justified. A brief discussion of its sensitivity (e.g., results with thresholds 40% and 60%) would strengthen the ensemble method.","section":"Eq. (1)"},{"comment":"The embedding dimension sweep is described as 'between 2 and min(33, D) in increments of 3'; please clarify whether the grid is {2,5,8,...,32} or includes other values, and how Rmax=33 was chosen.","section":"§5.1"},{"comment":"The class ratio column mixes notations such as '731:1.1' and '91:1.9'; this is confusing because the majority/minority ordering is not consistent. Use a single convention (e.g., majority:minority) throughout.","section":"Table 1"},{"comment":"No code or configuration details are provided for reproducing the experiments. Since the paper is empirical, releasing code and hyperparameter settings would substantially improve reproducibility.","section":"General"},{"comment":"The limitation that the ensembles handle only binary clustering is stated, but the paper does not discuss how the label-alignment step (Eq. 1) would be extended to multiple classes. This is a known limitation, not a blocker, but a sentence on a potential multi-class alignment strategy would be helpful.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The core empirical finding that ensembles can outperform individual clustering methods on these datasets is credible, but the paper's most prominent claim—that deep clustering is a 'strong alternative to supervised classification'—is outside the evidence presented. The post hoc selection of KGT's members is a second load-bearing issue. I would not reject because both issues are fixable within the manuscript's scope: adding supervised baselines and a validation-based selection protocol. I suggest the editor require these changes before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take.\n\nThe paper is worth reading for the comparison, not for the headline. It puts six existing deep clustering methods on 16 binary tabular datasets under controlled imbalance and adds two ensemble strategies. The systematic benchmark with induced 1:1 to 1:15 ratios is genuinely new and useful, and the G-CEALS-EE idea—aligning cluster labels across embedding dimensions and averaging—is simple, sensible, and looks like a real improvement over G-CEALS, which is known to be sensitive to embedding dimension. KGT is a majority vote of three methods, but it has a selection problem.\n\nThe soft spots are real. First, the abstract claims deep clustering 'can serve as a strong alternative to supervised classification,' yet no supervised classifier is run anywhere. That claim is untested. Fixing other issues would not fix this; the authors need logistic regression, XGBoost, or a standard classifier on the same splits. Second, KGT is built by ranking methods on the same 16 datasets used for evaluation and taking the top three. That is selection on the test set. The paper states labels are not used for hyperparameter optimization, but member selection uses the evaluation metrics on those same datasets, so that statement is misleading. G-CEALS-EE does not have this flaw and stands on its own. Third, there are no significance tests. Table 2 shows min-max normalized scores averaged within groups; error bars in Figure 1 are large, and without paired tests or absolute metric values, the differences could be noise. Fourth, no code is released for the ensemble methods, which makes replication harder.\n\nOn the positive side, the authors are clear about the binary-only limitation and do not overclaim elsewhere. The experimental setup—natural imbalance groups plus controlled downsampling—is a reasonable way to separate dataset-specific effects from the imbalance signal. The references are relevant. This is not a sloppy paper; it is an honest benchmark that overshoots in its conclusion.\n\nWho is this for? Researchers working on tabular clustering and representation learning, especially those wanting a quick map of which methods behave well under imbalance. I would bring it to a reading group because it will generate a useful discussion about evaluation standards. I would not cite it as evidence that deep clustering rivals supervised classification; I might cite it as a benchmark if the authors release code and rerun with the fixes. Peer review: this deserves a serious referee, and a good one will send it back for major revision rather than desk reject.","headline":"Useful first benchmark of deep clustering on imbalanced tabular data with one clever ensemble, but the paper overclaims a supervised comparison and pokes a hole in its own KGT selection.","tokens_in":10222,"tokens_out":3481,"would_cite":true,"duration_ms":31433,"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":"Unsupervised cluster ensembles recover true classes from imbalanced tabular data more reliably than single methods.","keywords":["deep clustering","cluster ensemble","class imbalance","tabular data","G-CEALS-EE","KGT ensemble","unsupervised learning","embedding dimensions"],"falsifier":"Split the 16 datasets into two halves: rank the individual methods on one half, build KGT from the top three found there, evaluate both KGT and the single methods on the other half, and repeat across many splits. If KGT no longer beats the best individual method on average, the reported ensemble resilience is a test-set selection artifact.","tokens_in":9415,"feed_emoji":"📊","tokens_out":8000,"duration_ms":61445,"temperature":0.7,"pith_summary":"Class imbalance biases supervised classifiers toward the majority class, so this paper asks whether clustering without labels can avoid the problem. It compares six existing clustering methods and two new ensembles on 16 binary tabular datasets under natural and artificially induced imbalance, from 1:1 to 1:15. The central claim is that the ensembles—G-CEALS-EE, which aligns and averages cluster assignments across autoencoder embedding dimensions, and KGT, which majority-votes K-means, G-CEALS-EE, and TabClustPFN—achieve the best average performance in accuracy, normalized mutual information, and adjusted Rand index. If this is right, deep clustering offers a label-free route to recovering minority classes in imbalanced tabular data, without the sacrifices of undersampling or the distortions of oversampling.","feed_headline":"Ensembles beat single clustering on imbalanced tabular data","feed_subtitle":"Voting across K-means, TabClustPFN, and an embedding ensemble recovers labels under 1:1 to 1:15 imbalance","key_machinery":"The alignment rule of Equation 1 is the load-bearing piece: cluster labels are arbitrary numbers, so before any aggregation the binary assignments from each embedding (for G-CEALS-EE) or each method (for KGT) are compared with a reference—the r=2 embedding or K-means—and flipped unless they agree on at least half the samples. This turns incomparable labelings into a common vote. Averaging the aligned binary assignments across dimensions and thresholding at 0.5 yields the ensemble label; KGT replaces averaging with majority voting across three aligned methods.","core_discovery":"The paper claims that combining cluster assignments from several unsupervised views outperforms any single clustering method under class imbalance. G-CEALS-EE varies the embedding dimension of the Gaussian-cluster autoencoder G-CEALS, treats the two-dimensional embedding as the reference, flips each other embedding's binary assignments whenever they match the reference on fewer than half the samples, then averages the aligned assignments and thresholds at 0.5. KGT takes K-means, G-CEALS-EE, and TabClustPFN, aligns their assignments against K-means in the same way, and labels each sample by majority vote. On the 16 datasets, both ensembles tie for the best overall rank and prove the most stab","pith_inferences":["The 50% flip rule depends on having exactly two clusters; extending to multiclass would need a permutation-alignment strategy such as Hungarian matching across runs, a natural next step.","Because KGT's members are chosen by ranking on the same 16 datasets used for evaluation, the reported advantage may partly reflect selection on the test set; ranking members on a validation split would test how much of the ensemble benefit generalizes.","The alignment-plus-average trick is not tied to G-CEALS specifically: any autoencoder-style method that produces clusterings at several latent dimensions could adopt the same recipe, so the approach could raise the imbalanced-data ceiling of other deep clusterers.","If the result holds, clustering ensembles could serve as a label-efficient first pass in domains with scarce labels—medical records, fraud detection—where minority classes are exactly the ones supervised models miss."],"forward_implications":["On the 16 binary tabular datasets, G-CEALS-EE and KGT tie for the best overall average rank across ACC, NMI, and ARI, ahead of every individual clustering method.","G-CEALS-EE is the most stable method under very high imbalance, taking best NMI and ARI and second-best ACC in that group, while TabClustPFN is the most imbalance-resilient individual method for NMI.","The ensembles deliver these results without labels, so they avoid class-weight tuning and do not require discarding majority samples or synthesizing minority samples.","Individual methods have complementary weaknesses—TabClustPFN's ACC is poor, QGRL degrades under imbalance, DCSS optimizes silhouette but not label agreement—so ensembling compensates rather than relying on any one objective."],"fun_headline_variants":["Clustering ensembles beat singles on imbalanced data","Ensemble deep clustering thrives on imbalanced tables","Majority voting across clusters handles data imbalance","Deep clustering ensembles outperform single algorithms","Unsupervised ensemble clustering for skewed tabular data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The KGT ensemble's edge depends on picking its three members using the same 16 datasets that later measure its performance; if the member choice were made on separate data, the ensemble advantage could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Clustering ensembles beat singles on imbalanced data","Ensemble deep clustering thrives on imbalanced tables","Majority voting across clusters handles data imbalance","Deep clustering ensembles outperform single algorithms","Unsupervised ensemble clustering for skewed tabular data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1250,"prompt_tokens":726,"completion_tokens":524,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":470,"tokens_out":524,"duration_ms":5211,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:38:56.385662+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Split the 16 datasets into two halves: rank the individual methods on one half, build KGT from the top three found there, evaluate both KGT and the single methods on the other half, and repeat across many splits. If KGT no longer beats the best individual method on average, the reported ensemble resilience is a test-set selection artifact.","supporting_citations":[],"review_version":1}