{"id":"0f55a194-bf56-4d68-9d45-40581728267a","arxiv_id":"2506.05877","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ICE builds an interpretable decision tree for clustering ensemble by splitting on original features to maximize summed chi-squared agreement with base k-means partitions.","lead":"Clustering ensembles combine many clusterings into one, but the final result is usually a black box. This paper builds the combination as a decision tree over original features, so each cluster assignment can be read as a simple set of rules.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section III-C1's sum-of-chi-squared p-values assume independence among dependent k-means partitions, so the invalid p-values control node-growth order and the statistical grounding of ICE is not established.","rationale":"The central claim is that ICE is the first interpretable clustering ensemble with quality comparable to state-of-the-art methods. The algorithm's novelty hinges on using statistical association tests to build a decision tree, and the reader's weakest assumption correctly identifies the independence assumption in Section III-C1. I agree with that diagnosis: the base partitions are k-means clusterings of the same data, so the chi-squared statistics are not independent, and the sum is not chi-squared distributed. This is the single most load-bearing concern because the p-value ranking is the sole mechanism for selecting which node to split next (Algorithm 1, line 10); if the ranking is miscalibrated, the tree's growth order is fundamentally unprincipled, and the paper's 'statistical association test' framing collapses. I would add two refinements. First, the expected-cell-count condition for the chi-squared approximation is not enforced, which further undermines even the within-node use of the statistic, though the within-node split choice uses the raw sum rather than the p-value. Second, the paper's own limitation statement in Section V acknowledges that p-value-based selection may not be optimal, which corroborates the concern. The threat is not necessarily fatal: the method could be reframed as a heuristic and still be accepted as an interpretable ensemble. But the current manuscript's novelty claim, which emphasizes statistical testing, is not sound as written. A permutation test would settle whether the p-values are informative; if they are not, the authors must either correct the calibration or explicitly present the criterion as a heuristic. The reader's conditional verdict is appropriate, and no code release further impedes independent verification, so the conditional status should remain unchanged.","tokens_in":20289,"tokens_out":5317,"duration_ms":55278,"concrete_test":"Run ICE on a null dataset (e.g., permuted features or permuted base-partition labels, keeping marginals) for each of several UCI sets; if p-values were valid, the minimum p-value across candidate nodes should be approximately uniform during early growth. Instead, compute the empirical distribution of the summed statistic under permutation of base-partition labels within each node. If the observed p-values deviate strongly from the permutation null, and if the node expansion order obtained with permutation-calibrated p-values yields materially different trees than the paper's chi-squared p-values, then the statistical grounding is load-bearing and must be replaced by a heuristic or a corrected test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C1 (Eq. 5) sums chi-squared statistics over c base partitions and invokes [39] to treat the sum as chi-squared with the summed degrees of freedom. This requires the c statistics to be independent. But Section IV-B generates every base partition by k-means on the same standardized data with cluster counts in [k,3k]; these partitions are strongly dependent, often agreeing on large blocks of samples. The resulting p-values are therefore miscalibrated. Algorithm 1, line 10, uses only these p-values to choose which candidate node to expand, so the greedy topology of the tree rests on invalid significance figures. Additionally, the chi-squared approximation requires expected cell counts at least 5, which is not checked; with c=30 and up to 3k categories, many cells are sparse, especially after splitting. The paper itself concedes in Section V that p-value-based selection 'may not always yield the most optimal partition' and has 'inherent limitations.' The concern is not that the algorithm fails empirically; it may work as a heuristic. The concern is that the paper's central justification—a statistical association test guiding the tree—is not statistically valid, and without a valid test the novelty claim of a principled interpretable clustering ensemble is unsupported. A permutation-based calibration, or explicitly relabeling the method as heuristic, is required.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICE, a decision-tree-based clustering ensemble. Base partitions are obtained by k-means; each candidate split is scored by summing the chi-squared statistic between the binary split and each base partition, and the sum is converted to a p-value under a chi-squared assumption. ICE grows the tree by repeatedly expanding the candidate node with the smallest p-value until k leaves are produced, yielding an interpretable tree in the original feature space. Experiments on 33 UCI/KEEL datasets compare ICE with k-means, two interpretable tree-based clustering methods, and several clustering ensemble baselines using Purity, F1-score, NMI, tree depth, and runtime. The paper claims that this is the first interpretable clustering ensemble algorithm and that it achieves performance comparable to state-of-the-art ensemble methods.","tokens_in":20577,"tokens_out":5192,"duration_ms":53702,"significance":"The idea of combining clustering ensembles with decision-tree interpretability is timely and, if the statistical criterion were valid, ICE would be a useful contribution. The algorithm is concrete, avoids fitting parameters to ground-truth labels, and the incremental contingency-table update is a practical implementation detail with a plausible time-complexity analysis. The experimental study is broad, covering 33 datasets. However, the validity of the p-value criterion is load-bearing: it determines both the split chosen inside a node and the order in which nodes are expanded, so the entire statistical justification of the tree-building process rests on an assumption that is not satisfied. The empirical comparison also lacks significance testing. These issues require substantive rework before the central claims can be accepted.","major_comments":[{"comment":"The sum in Eq. (5) is treated as chi-squared with summed degrees of freedom, citing [39], but this is valid only if the c chi-squared statistics are independent. Section IV-B generates every base partition by applying k-means to the same standardized data, so the base partitions, and hence the statistics chi^2(S_j, pi_t), are strongly dependent. The p-values used in Algorithm 1 line 10 are therefore miscalibrated, and comparing p-values across nodes with different sample sizes and different numbers of base-partition categories compounds the problem. Since this p-value determines which node is expanded next, the global shape of the tree rests on an unjustified distributional assumption. The limitation noted in Section V, that p-value selection 'may not always yield the most optimal partition', does not address this validity problem. Please replace the asymptotic p-value with a permutation-based or other calibration that preserves the dependence structure, or explicitly reframe ICE as a heuristic and remove the statistical-test claim.","section":"Section III-C1 (Eq. 5) and Algorithm 1 (line 10)"},{"comment":"Even setting aside the dependence issue, the chi-squared approximation requires expected cell counts of at least about 5 in the contingency table. Algorithm 1 only checks that each child node has at least 5 samples, not that each cell's expected count is adequate. With c=30 and base-partition cluster counts up to 3k, many cells are sparse after splits, especially in smaller nodes. This additional miscalibration should be addressed in any revision that retains the statistical-test framing.","section":"Section III-C1 and Algorithm 1 (lines 19-23)"},{"comment":"The claim of 'comparable performance' to state-of-the-art ensemble methods is not supported by any significance test, such as a Wilcoxon signed-rank test across the 33 datasets. Reading the averages, ICE has lower mean NMI than every ensemble baseline (0.2852 versus 0.3074-0.3119) and lower mean Purity than most ensemble baselines. The paper should either provide a proper paired comparison with per-dataset wins/losses and a significance test, or soften the claim to 'competitive on some datasets'.","section":"Section IV-C (Tables III-V)"}],"minor_comments":[{"comment":"Section IV-A says the comparison uses 34 datasets, while Section IV-B and Table II list 33 datasets; please reconcile this discrepancy.","section":"Section IV-A and IV-B"},{"comment":"There are several typographical errors, including 'statisitc' in Algorithm 1 and 'recieved' in the manuscript header; these should be corrected.","section":"Algorithm 1 and manuscript header"},{"comment":"The statement that the sum in Eq. (5) follows a chi-squared distribution should explicitly state the independence condition and cite the precise theorem, rather than a general categorical-data-analysis reference.","section":"Section III-C1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope, but the central statistical rationale is currently invalid. The authors should either provide a valid calibration of the split-quality measure or reframe the method as a heuristic; in either case, the experimental comparison needs significance testing. With those changes, the contribution could be acceptable, but as it stands the main claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously: treating each base partition as a categorical variable and growing a decision tree whose splits maximize the summed chi-squared association across the ensemble. That appears to be new, and the algorithm is concrete with a plausible complexity analysis. The authors extend their own significance-based tree clustering work in a sensible direction.\n\nThe problem is the statistical engine. In Section III-C1, they sum the chi-squared statistics over the c base partitions and treat the sum as chi-squared with summed degrees of freedom, citing Agresti. That requires independence. But the base partitions are all k-means runs on the same standardized data, so they are strongly dependent. The resulting p-values are miscalibrated, and Algorithm 1's greedy node selection rests on those invalid numbers. The authors themselves concede in Section V that the p-value strategy \"may not always yield the most optimal partition.\" This doesn't kill the method as a heuristic - it might still find reasonable splits - but the paper's central claim of principled statistical association driving the tree is unsupported. They need a permutation-based calibration or an explicit reframing as a heuristic score.\n\nThe experiments are also weaker than the abstract suggests. The averages in Tables III-V show ICE generally below the co-association and graph-based ensemble methods on purity and NMI; it's \"comparable\" only in a loose sense. There are no error bars or significance tests, and no code is released. The interpretability comparison against IMM and SHA is fine but minor.\n\nThe citation pattern is fine, and the method doesn't fit any parameters to ground-truth labels, which is a plus. Overall, this deserves a serious referee, but only with major revision: fix or reframe the statistical step, add proper experimental comparisons with variance, and release code. As it stands, the claims outrun the evidence.","headline":"A genuinely new idea for an interpretable clustering ensemble, undermined by an invalid p-value calculation and overclaimed experimental support.","tokens_in":575,"tokens_out":951,"would_cite":true,"duration_ms":29381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Clustering ensembles get a decision tree you can read, without losing accuracy.","keywords":["clustering ensemble","interpretable clustering","decision tree","chi-squared test","consensus partition","unsupervised learning","p-value","k-means"],"falsifier":"On a dataset with known ground-truth clusters, compute the average absolute correlation between the c base partitions at the root node; if it is high (say above 0.5), the independence assumption is violated at the very first split. A more direct test is to run ICE twice, once with the paper's p-value and once with a permutation-based p-value that breaks dependence, and compare the order in which nodes are split and the final purity: if the two orders diverge materially on a majority of datasets, the independence-based p-value is not a reliable guide for tree growth.","tokens_in":20100,"feed_emoji":"🌳","tokens_out":5901,"duration_ms":54415,"temperature":0.7,"pith_summary":"Clustering ensembles combine many clusterings into one consensus partition, but the result is usually a black-box label assignment. This paper proposes ICE, the first clustering ensemble method that outputs a decision tree in the original feature space, so every cluster assignment can be traced along splits like \"feature ≤ threshold\". At each growth step, ICE picks the split whose association with the base partitions—measured by summed chi-squared statistics converted to a p-value—is strongest, and keeps growing until it reaches the requested number of clusters. The authors claim this yields clustering quality comparable to state-of-the-art ensemble methods while adding interpretability, and outperforms existing interpretable clustering methods in accuracy. If true, high-stakes applications like medical diagnosis and financial risk assessment could use ensemble stability without giving up explanations.","feed_headline":"First clustering ensemble that explains itself as a decision tree","feed_subtitle":"ICE matches state-of-the-art ensemble accuracy while every cluster is traceable to a threshold split on original features.","key_machinery":"The central mechanism is the split-scoring statistic: for a candidate split $S_j$ defined as $f_j \\le v_j$, the algorithm computes the chi-squared statistic $\\chi^2(S_j, \\pi_t)$ between the split and each base partition $\\pi_t$, then sums them as $\\chi^2(S_j) = \\sum_{t=1}^c \\chi^2(S_j, \\pi_t)$. Treating this sum as a chi-squared random variable under the independence assumption, it converts the score to a p-value, which allows splits from different nodes with different sample sizes and different numbers of base-partition categories to be compared directly. The greedy growth rule—always split the candidate node with the smallest p-value—is what ties the statistical test to the tree-construction process and makes each split an explainable, ensemble-driven decision.","core_discovery":"ICE treats each base partition as a categorical variable and builds a binary decision tree whose splits are predicates of the form feature ≤ threshold. For every candidate split at a node, it builds a contingency table of the split against each base partition, computes the chi-squared statistic per table, sums these statistics, and converts the sum to a p-value under the assumption that the c chi-squared components are independent. The split with the smallest p-value becomes the node's best split, and the algorithm grows the tree by splitting, at each round, the candidate node whose best split has the smallest p-value, until k leaves exist. The returned tree is itself the consensus partition: each leaf is a cluster, and each internal node explains in original features why the data was separated. The paper reports experiments on 33 benchmark datasets showing average Purity, F1-score, and NMI comparable to state-of-the-art clustering ensemble methods and better than the interpretable baselines IMM and SHA, with tree depths comparable to or smaller than IMM's.","pith_inferences":["The statistical test is being used as a model-selection score rather than a hypothesis test; because the base partitions are dependent, the p-values are likely miscalibrated, and a more honest reading is that the chi-squared sum is a useful heuristic score. A permutation-based or corrected significance test could preserve the interpretability while making the ranking valid.","The same split-scoring idea could be lifted to other black-box clusterings (e.g., spectral or DBSCAN) as long as each base partition can be expressed as a categorical label, turning ICE into a general interpreter for any ensemble.","For high-dimensional data, the exhaustive threshold search will dominate runtime; one testable extension is to restrict candidate thresholds to quantiles or to use a surrogate score to prune features before the chi-squared computation.","If the goal is explanation fidelity rather than consensus quality, ICE's tree could be compared against a decision-tree approximation of the consensus labels from any ensemble method; the split-by-association criterion directly optimizes agreement with base partitions, which may make it a stronger explanation of the ensemble than a post-hoc fit."],"forward_implications":["Clustering ensemble results become inspectable: each final cluster corresponds to a leaf, and its defining conditions are simple threshold rules on original features, so a domain expert can verify whether a consensus cluster is meaningful.","Because the split choice is driven by association with the ensemble rather than by class labels, the method applies to fully unlabeled data and can serve as a drop-in replacement for consensus functions in existing ensemble pipelines.","The fixed leaf count $k$ lets a user control the granularity of the explanation: smaller $k$ gives shallower, more readable trees, at some cost in accuracy.","With ensemble size $c$, clustering quality generally improves, and ICE's accuracy tracks that improvement, so the interpretability does not appear to cap the ensemble's ability to gain from more base clusterings.","The reported average tree depth (avgDepth 1.98, maxDepth 2.62) suggests ICE is competitive in compactness with dedicated interpretable clustering baselines, not just in accuracy."],"supporting_citations":[{"why":"Supplies the property that a sum of independent chi-squared variables follows a chi-squared distribution, which justifies converting the summed statistic into a p-value.","marker":"[39]"},{"why":"The IMM explainable k-means baseline that ICE is compared against for clustering accuracy and tree compactness.","marker":"[31]"},{"why":"The SHA shallow decision tree baseline for explainable k-means, used in accuracy and depth comparisons.","marker":"[7]"},{"why":"The LWEA and LWGP locally weighted clustering ensemble methods that serve as state-of-the-art quality baselines.","marker":"[40]"},{"why":"The ECCMS co-association matrix self-enhancement ensemble method used as a state-of-the-art baseline.","marker":"[41]"},{"why":"The ECPCS-HC and ECPCS-MC cluster-wise similarity propagation ensemble methods used as state-of-the-art baselines.","marker":"[42]"}],"fun_headline_variants":["Decision-tree clustering ensemble that's first to be interpretable","Interpretable ensemble clusters via feature thresholds","Clustering ensemble with a built-in decision tree explanation","First interpretable clustering ensemble—matches SOTA accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the c chi-squared statistics for a candidate split are independent, so their sum follows a chi-squared distribution and the resulting p-values rank the splits correctly; the base partitions are k-means runs on the same data, so they are strongly dependent, and if the p-value ranking is not meaningful the greedy choice of which node to split loses its statistical justification.","fun_headline_variants_meta":{"raw":{"variants":["Decision-tree clustering ensemble that's first to be interpretable","Interpretable ensemble clusters via feature thresholds","Clustering ensemble with a built-in decision tree explanation","First interpretable clustering ensemble—matches SOTA accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1209,"prompt_tokens":904,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":520,"tokens_out":305,"duration_ms":3479,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:13:06.815212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with known ground-truth clusters, compute the average absolute correlation between the c base partitions at the root node; if it is high (say above 0.5), the independence assumption is violated at the very first split. A more direct test is to run ICE twice, once with the paper's p-value and once with a permutation-based p-value that breaks dependence, and compare the order in which nodes are split and the final purity: if the two orders diverge materially on a majority of datasets, the independence-based p-value is not a reliable guide for tree growth.","supporting_citations":[{"cited_title":"Agresti,Categorical data analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the property that a sum of independent chi-squared variables follows a chi-squared distribution, which justifies converting the summed statistic into a p-value."},{"cited_title":"Explainable k-means and k-medians clustering,","cited_arxiv_id":null,"evidence_quote":"The IMM explainable k-means baseline that ICE is compared against for clustering accuracy and tree compactness."},{"cited_title":"Shallow decision trees for ex- plainable k-means clustering,","cited_arxiv_id":null,"evidence_quote":"The SHA shallow decision tree baseline for explainable k-means, used in accuracy and depth comparisons."},{"cited_title":"Locally weighted ensemble clustering,","cited_arxiv_id":null,"evidence_quote":"The LWEA and LWGP locally weighted clustering ensemble methods that serve as state-of-the-art quality baselines."},{"cited_title":"Ensemble clustering via co- association matrix self-enhancement,","cited_arxiv_id":null,"evidence_quote":"The ECCMS co-association matrix self-enhancement ensemble method used as a state-of-the-art baseline."},{"cited_title":"Enhanced ensemble clustering via fast propagation of cluster-wise similarities,","cited_arxiv_id":null,"evidence_quote":"The ECPCS-HC and ECPCS-MC cluster-wise similarity propagation ensemble methods used as state-of-the-art baselines."}],"review_version":1}