{"id":"735d4c41-fde7-47e6-ad8f-11f82cade147","arxiv_id":"2411.14728","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"K-GBS3FCM dynamically sizes each labeled point's unlabeled neighborhood and simplifies the safety-degree formula, reporting wins in 36 of 56 accuracy comparisons.","lead":"The paper presents a semi-supervised fuzzy clustering method that uses K-nearest-neighbor graphs to decide how much to trust labels when some are wrong. It reports higher accuracy than existing methods on most benchmark configurations, but the comparison relies on previously published baseline numbers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 64% superiority claim is not established: baselines from [17] were not rerun, so the comparison lacks matched splits, mislabeling masks, and statistical testing.","rationale":"The reader's REJECT verdict is appropriate. I considered whether the more serious problem is the algorithm's internal construction (e.g., Eq. 13 can produce non-integer dpUN when UNmax - UNmin is not integer; the Step(.) multiplier in Eq. 17 is not clearly specified in the parameter list; and no convergence proof is given for the alternating updates). These are real weaknesses, but they do not strike at the central claim as directly as the comparison protocol. The method could be correct and still fail to demonstrate superiority. Conversely, if the comparison were properly matched, the 64% claim could be checked. Since the paper explicitly says the baselines were not rerun, every win/loss comparison is confounded by the unknown equivalence of experimental setups. This is a single, falsifiable methodological flaw, and the proposed concrete test would settle it. If the authors rerun baselines under matched splits and the result still stands, the central claim would be supported (though error bars should also be reported). Until then, the correct verdict is reject as an empirical-comparison paper.","tokens_in":13851,"tokens_out":5099,"duration_ms":50431,"concrete_test":"Rerun all eight baseline algorithms (K-Means, FCM, SMKFCM, SMKFC-ER, CSNMF, LHCS3FCM, CS3FCM, AS3FCM) and K-GBS3FCM on the identical 20 random train/mislabeling splits per dataset and mislabeling rate, using the authors' public code (and baseline implementations from [17] or original papers). For each of the 56 cells, compute mean and standard deviation of CA over the 20 runs and run a paired Wilcoxon signed-rank test comparing K-GBS3FCM against the best baseline. If K-GBS3FCM does not win the same 36+ cells, or if the wins are not statistically significant at the cell level, the claimed 64% superiority is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim — that K-GBS3FCM is 'significantly superior' in 64% of 56 configurations — depends on comparing the authors' own runs against accuracy numbers taken from the AS3FCM paper. The paper states in Section IV-B: 'These algorithms have not been rerun; instead, the results published in [17] are relied upon.' This makes the comparison non-matched: the baselines may have used different random 20/80 splits, different mislabeling injections, different hyperparameter choices, and possibly a different accuracy computation (e.g., label matching). No paired error bars or significance tests are reported for the 20 runs per configuration, so the word 'significantly' is unsupported. Because the headline result is a comparative claim, the missing matched-baseline protocol is load-bearing: if the baseline numbers were produced under any different condition, the 64% win rate is not a valid estimate of relative performance. The algorithmic idea may still be coherent, but the paper as written does not substantiate superiority.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes K-GBS3FCM, a safe semi-supervised fuzzy c-means algorithm that extends AS3FCM. The proposed modifications are: (i) a dynamic number of unlabeled neighbors per labeled point, computed via an average KNN-distance density proxy; (ii) a simplified safety-degree estimate defined as the inverse of a normalized local inconsistency; and (iii) a step-function-activated amplification of the λ2 term when the average safety degree exceeds a threshold. The authors evaluate the method on eight datasets with mislabeling rates from 0% to 30%, comparing against eight baselines, and claim the method is significantly superior in 64% of the 56 test configurations.","tokens_in":14128,"tokens_out":7515,"duration_ms":68612,"significance":"The algorithmic ideas are coherent and potentially of interest to the safe semi-supervised fuzzy clustering community, and the authors provide publicly available code, which is a concrete strength. If the dynamic-neighborhood mechanism were validated under a sound experimental protocol, it would be a useful contribution. However, the current empirical evidence does not support the headline comparative claim: the baselines were not rerun under matched conditions, the hyperparameter selection is not validation-based, and no statistical significance testing is reported. The contribution is therefore not yet established.","major_comments":[{"comment":"The benchmark comparison is invalid as a test of superiority because the eight baselines were not rerun under the same protocol. The paper states: 'These algorithms have not been rerun; instead, the results published in [17] are relied upon.' Consequently, the baseline accuracy values may come from different random 20/80 splits, different mislabeling injection masks, different hyperparameter tuning, and possibly different evaluation code. The abstract's claim of 'significantly superior in 64% of the 56 test configurations' is therefore not supported. The authors should rerun all baselines on the same splits, with the same mislabeling procedure, and report paired error bars and significance tests.","section":"Section IV-B"},{"comment":"The hyperparameter selection procedure is under-specified. λ1 and λ2 are searched over a 6×6 grid, but the paper does not state whether a validation split was used to choose them. If the reported accuracies (Table II) are the best test-set values over the grid, the results are positively biased and not comparable to baselines tuned differently. The authors must describe the selection rule and use a validation-based procedure for all methods.","section":"Section IV-C"},{"comment":"Eqs. (17)-(19) contain an undefined multiplier written as 'λ2λStep(¯s) 1'. The surrounding text says the step value 'activates a multiplier of λ2 part', while Algorithm 1 says the λ2 part is 'amplified by λ1'. This ambiguity affects the core objective function and the membership update equations. The authors should clarify the correct form and confirm that (18) and (19) are exact derivatives of (17).","section":"Section III-C"},{"comment":"The algorithm is an iterative reweighting heuristic rather than a solver for a single objective function: sk in Eq. (15) is computed from the current memberships uir, while those memberships are updated in Eqs. (18)-(19) using the previous sk. This is not necessarily wrong, but the paper should explicitly state that the objective (17) is minimized only over u and v with sk held fixed, and discuss the convergence behavior. In addition, dpUNk in Eq. (13) is a real number but is used to select a number of nearest neighbors; the rounding scheme is not specified.","section":"Section III-B and Algorithm 1"},{"comment":"The term 'significantly superior' is not supported by any statistical evidence. The paper reports 20 runs per configuration but provides no standard deviations, confidence intervals, or paired significance tests. Even if the baselines had been rerun, the claim of significance would require such tests. The authors should report these statistics and state the exact test used.","section":"Section IV-E"}],"minor_comments":[{"comment":"The text says 'The results of these five algorithms' but eight algorithms are listed; this should be corrected.","section":"Section IV-B"},{"comment":"Please specify how dpUNk is rounded to an integer before it is used as a KNN parameter.","section":"Eq. (13)"},{"comment":"Please describe how predicted labels are aligned with true labels (e.g., Hungarian matching); without this, the accuracy definition is ambiguous.","section":"Eq. (20)"},{"comment":"The sentence 'the λ2 part of the objective function is amplified by λ1' appears to be a typo; it should probably refer to Step(¯s) as the multiplier.","section":"Section III-C"},{"comment":"The complexity discussion does not justify the claim that K-GBS3FCM 'would typically perform better' than AS3FCM; this assertion should be supported with a comparative complexity table or removed.","section":"Section III-E"}],"recommendation":"major_revision","confidential_remarks":"The paper's central comparative claim rests on a protocol that cannot support it, but the algorithm itself is coherent and the code is available. I would encourage the editor to invite a thorough experimental revision rather than reject outright. The authors should also be asked to clarify the objective function and update equations before any further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: K-GBS3FCM is a coherent, clearly described incremental extension of AS3FCM, but the headline empirical claim — 'significantly superior in 64% of 56 test configurations' — is not supported by the reported methodology. The comparison against baselines is not matched: the authors took the baseline numbers from [17] without rerunning them, so the random splits, mislabeling masks, hyperparameter choices, and even the accuracy computation could differ. That alone makes 'significantly' unsupported.\n\nWhat's actually new: the dynamic per-point unlabeled neighbor count (dpUN), the threshold-activated amplification of the λ2 term, and a simplified closed-form safety degree derived from membership consistency in (15). These are genuine, modest changes. The dynamic neighbor idea addresses a real limitation of AS3FCM's fixed pUN, and the simplification is reasonable. The authors also drop the sum-to-one constraint on safety degrees, which is worth testing. The failure on Gauss50x is acknowledged and attributed to cluster overlap, which is honest. Code and results are on GitHub, which I appreciate.\n\nSoft spots: the main one is the load-bearing comparison. Section IV-B explicitly says the baselines were not rerun. That is a real problem: the reported accuracy of K-GBS3FCM is produced under the authors' own hyperparameter sweep over 36 λ combinations with no validation split, so it is likely optimistic. Without statistical tests or error bars, we cannot know whether the 64% win rate is signal or artifact. The circularity of computing safety degrees from memberships that are then reweighted by those safety degrees is present, but it is a minor concern here because s is not optimized in the objective.\n\nBottom line: for a reader in safe semi-supervised fuzzy clustering, this is a competent incremental paper with a plausible algorithmic idea. It deserves a serious referee, but not as is. The authors need to rerun the baselines under identical splits and mislabeling protocols, add a validation split for hyperparameters, and report error bars or significance. If they do that, the empirical claim might hold. I would not cite this version in my own work, and I would not prioritize it for a reading group.","headline":"A coherent incremental variant of AS3FCM with a promising dynamic-neighborhood idea, but the 64% superiority claim is unproven because baselines were not rerun under matched conditions.","tokens_in":14547,"tokens_out":3724,"would_cite":false,"duration_ms":34076,"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":"The paper claims K-GBS3FCM, a fuzzy c-means variant that sizes each labeled point's unlabeled neighborhood by KNN density and boosts the labeled-to-unlabeled influence when average safety is high, wins 36 of 56 benchmark configurations…","keywords":["safe semi-supervised clustering","fuzzy c-means","K-nearest neighbors graph","safety degree","mislabeled data","semi-supervised fuzzy clustering","clustering accuracy","label noise"],"falsifier":"A direct check is to rerun the eight baseline algorithms on the same eight datasets under the paper's stated protocol (20% labeled, 0-30% mislabeling in 5% steps, accuracy metric of Eq. 20) and see whether K-GBS3FCM still wins 36 of 56 configurations; if the baselines reproduce the [17] numbers and the win rate drops well below 64%, the central claim fails, and if the baselines do not reproduce, the original comparison is not trustworthy.","tokens_in":13679,"feed_emoji":"📊","tokens_out":6316,"duration_ms":53993,"temperature":0.7,"pith_summary":"The paper introduces K-GBS3FCM, a safe semi-supervised fuzzy c-means algorithm that uses a K-nearest-neighbor graph to decide how many unlabeled neighbors each labeled point should consult. This is meant to fix a weakness in the prior AS3FCM method, which fixed that neighbor count at 5 regardless of local density. The paper claims the new algorithm is significantly better than eight baselines in 64% of 56 configurations across eight datasets, including cases with up to 30% intentionally wrong labels. If true, it supports the idea that graph-based neighborhood estimation plus a simple reliability threshold can make semi-supervised fuzzy clustering safer under label noise.","feed_headline":"KNN graph helps safe fuzzy clustering win 64% of benchmark tests","feed_subtitle":"Dynamic neighborhoods and a safety threshold beat semi-supervised and unsupervised baselines under label noise.","key_machinery":"The central object is the KNN-based dynamic neighborhood size dpUN_k (Eq. 13), which maps the average distance to K nearest unlabeled neighbors to a value between UN_min and UN_max. This dpUN feeds a simplified safety degree s_k = 1/(1 + local_inconsistency/dpUN) (Eq. 15), and a step function Step(mean s) (Eq. 16) multiplies the λ2 regularization term in the objective function when the average safety degree is reliable. The machinery ties the density proxy to how much each labeled point is trusted and how strongly it influences nearby unlabeled points.","core_discovery":"K-GBS3FCM replaces the fixed number of unlabeled neighbors used in AS3FCM with a per-labeled-point dynamic number dpUN, estimated from the average distance to a fixed K initial neighbors as a density proxy. It simplifies the safety-degree computation to a closed form based on local inconsistency, and it adds a step function that amplifies the influence of labeled data on unlabeled data when the average safety degree passes a threshold. The paper reports that this combination achieves higher clustering accuracy than FCM, K-Means, and several recent safe semi-supervised methods, and holds up better as mislabeling rises.","pith_inferences":["The step function is a hard switch: it either fully activates or fully deactivates the boosted λ2 term. A softer weighting could yield smoother behavior on datasets near the threshold, which is a testable extension.","Because dpUN is a linear rescaling of mean distance, it is sensitive to outliers that inflate the maximum distance. Using a robust density proxy or percentile scaling might improve performance on datasets like Gauss50x.","The simplification from AS3FCM's quadratic-programming safety degree to a closed-form consistency score suggests that the extra complexity may not be necessary; a head-to-head rerun of AS3FCM under identical conditions would test this directly."],"forward_implications":["With correct labels (0% mislabeling), K-GBS3FCM outperforms all baselines on all datasets except Gauss50x.","Through 30% mislabeling, K-GBS3FCM wins 36 of 56 configurations, i.e., 64.28% of the tested cases.","Accuracy degrades gracefully with mislabeling, with the worst drop being 9.2 percentage points on the Bupa dataset (from 69.3% to 60.1%).","The algorithm loses to other semi-supervised methods only on the heavily overlapping Gauss50x dataset, where KNN neighborhoods mix clusters and safety degrees become unreliable."],"supporting_citations":[{"why":"Supplies the AS3FCM method this work builds on, the objective-function structure, the eight benchmark datasets, and the published baseline results used for comparison.","marker":"[17]"},{"why":"Defines SSFCM, the semi-supervised fuzzy c-means with a fidelity term that the proposed objective function extends.","marker":"[8]"},{"why":"Provides the original FCM objective function and iterative membership/center update scheme underlying the algorithm.","marker":"[3]"},{"why":"Introduces the k-nearest-neighbor method used to compute the dynamic unlabeled-neighbor count and local consistency.","marker":"[20]"},{"why":"Introduces the safety-aware semi-supervised learning concept and safety-degree measure that the paper adapts to clustering.","marker":"[9]"}],"fun_headline_variants":["KNN graph boosts safe fuzzy clustering to 64% win rate","Safe fuzzy clustering with KNN graph beats baselines in 64% of tests","Dynamic neighbor safety in fuzzy C-means wins 64% of benchmarks","Graph-based safe fuzzy clustering: 64% better than other methods","K-GBS3FCM: KNN safety net for fuzzy clustering wins 64% of tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the published baseline numbers taken from AS3FCM [17] were produced under the same experimental conditions as the authors' own runs, since those baselines were not rerun (the paper states this in Section IV-B); if the random splits, tuning, or evaluation code differ, the claimed superiority rests on an apples-to-oranges comparison.","fun_headline_variants_meta":{"raw":{"variants":["KNN graph boosts safe fuzzy clustering to 64% win rate","Safe fuzzy clustering with KNN graph beats baselines in 64% of tests","Dynamic neighbor safety in fuzzy C-means wins 64% of benchmarks","Graph-based safe fuzzy clustering: 64% better than other methods","K-GBS3FCM: KNN safety net for fuzzy clustering wins 64% of tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001165,"raw_usage":{"total_tokens":4804,"prompt_tokens":912,"completion_tokens":3892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":3790}},"tokens_in":528,"tokens_out":3892,"duration_ms":26362,"temperature":1.0,"reasoning_tokens":3790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:58:07.409360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check is to rerun the eight baseline algorithms on the same eight datasets under the paper's stated protocol (20% labeled, 0-30% mislabeling in 5% steps, accuracy metric of Eq. 20) and see whether K-GBS3FCM still wins 36 of 56 configurations; if the baselines reproduce the [17] numbers and the win rate drops well below 64%, the central claim fails, and if the baselines do not reproduce, the original comparison is not trustworthy.","supporting_citations":[{"cited_title":"Adaptive safety-aware semi-supervised clustering,","cited_arxiv_id":null,"evidence_quote":"Supplies the AS3FCM method this work builds on, the objective-function structure, the eight benchmark datasets, and the published baseline results used for comparison."},{"cited_title":"Fuzzy clustering with partial supervision,","cited_arxiv_id":null,"evidence_quote":"Defines SSFCM, the semi-supervised fuzzy c-means with a fidelity term that the proposed objective function extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the original FCM objective function and iterative membership/center update scheme underlying the algorithm."},{"cited_title":"Fix, Discriminatory analysis: nonparametric discrimination, consis- tency properties","cited_arxiv_id":null,"evidence_quote":"Introduces the k-nearest-neighbor method used to compute the dynamic unlabeled-neighbor count and local consistency."},{"cited_title":"Safety-aware semi-supervised classification,","cited_arxiv_id":null,"evidence_quote":"Introduces the safety-aware semi-supervised learning concept and safety-degree measure that the paper adapts to clustering."}],"review_version":1}