{"id":"68a994c8-4eae-4e9c-b6a5-d0b2d92d6894","arxiv_id":"1908.04771","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors combine non-negative matrix factorization with fuzzy c-means to learn a shared hidden space and cluster it, reporting improved performance over comparison methods on six datasets.","lead":"This paper proposes a multi-view clustering method that learns a shared hidden representation of all views using non-negative matrix factorization while clustering that representation with fuzzy c-means. The method reports higher clustering accuracy than eight multi-view and two single-view clustering algorithms on six benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical superiority claim is weakened by test-set parameter selection: all methods, including HSS-MVFC with four free parameters, are tuned on the same datasets used for evaluation, so the reported rank advantage may be an artifact of test-set overfitting.","rationale":"The reader's weakest assumption concerns the weight update Eq. (20d). From the text as typeset, Eq. (20d) is the standard softmax solution for an entropy-regularized objective that is linear in w_k, so the claimed inconsistency depends on whether Eq. (15) really contains w_k^λ rather than w_k; this should be clarified but is not the decisive issue for the paper's empirical claim. The more load-bearing weakness is the evaluation protocol: all algorithms, including HSS-MVFC, are tuned on the test datasets via grid search, and no validation split or selected parameter values are reported. This can inflate the apparent advantage of a method with four tuned parameters and makes the central claim of broad superiority hard to verify independently. The Holm tests also show non-significant differences against several strong baselines, so the prose overstates the evidence. These concerns justify the reader's CONDITIONAL verdict rather than outright acceptance; code release or an honest validation study would be the natural way to resolve them.","tokens_in":16436,"tokens_out":16385,"duration_ms":176090,"concrete_test":"For each of the six datasets, perform repeated 5-fold cross-validation: tune λ, η, r, and m (and the baselines' parameters) on the training/validation folds only, fix the selected values, and report RI/NMI on the held-out test folds. Recompute mean ranks and the Friedman/Holm comparisons from these honest test-set results; if HSS-MVFC no longer ranks first, the headline superiority claim should be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HSS-MVFC clusters better than many related methods, and the evidence is Tables IV and V. Those tables report means over 10 runs at parameter values chosen from the grid-search ranges in Table II. The paper never describes a validation split or an independent model-selection procedure; for HSS-MVFC the grid includes λ, η, r, and m, four free parameters. Selecting them on the same test sets used to compute RI/NMI gives an optimistically biased comparison, and the bias is largest for the method with the most tuning freedom. The paper's own Holm post-hoc tests fail to reject equality with several baselines on both RI and NMI, so the support for 'better than many' is narrower than the abstract suggests. Without code or a report of the selected parameter values, an independent check cannot separate a genuine algorithmic advantage from test-set overfitting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HSS-MVFC, a multi-view fuzzy c-means variant in which each view is projected through a view-specific nonnegative basis matrix into a common low-dimensional hidden space H, a shared fuzzy partition U and cluster centers V are learned on H, and view weights w are learned by minimizing an entropy-regularized objective. The authors derive alternating updates for U, V, P_k, H, and w, give a complexity analysis, and compare against ten methods on six real datasets using RI and NMI together with Friedman and Holm tests. The central claim is that HSS-MVFC achieves better clustering performance than many related multi-view methods.","tokens_in":16635,"tokens_out":7886,"duration_ms":78338,"significance":"The modeling idea is coherent and potentially valuable: coupling NMF-based hidden-space discovery with fuzzy partition learning in a single objective is a natural extension of existing NMF-based multi-view clustering, and the maximum-entropy weighting is a sensible mechanism to avoid degenerate view weights. The paper provides explicit update rules and an experimental study with statistical tests on six datasets, which are strengths. However, the derivation of the weight update does not match the stated objective, and the experimental protocol appears to select parameters on the test sets, so the claimed superiority over baselines is not currently established. If the technical mismatch is resolved and the evaluation is made unbiased and reproducible, the method could be a useful contribution.","major_comments":[{"comment":"The closed-form weight update (20d) is not the minimizer of subproblem P5 for the objective stated in (15). The second term of (15) contains w_k^λ D_k, where D_k = ||X_k - P_k H||_F^2, but the Lagrangian in (20a) and the derivation leading to (20b)-(20d) treat that term as w_k D_k. With the exponent λ present, stationarity gives λ w_k^{λ-1} D_k + η(ln w_k + 1) + γ = 0, which does not have the closed form shown in (20d); the update in (20d) is exact only for λ = 1 and no exponent on D_k. Consequently the iterates optimize a different objective than (15) whenever λ ≠ 1, and the parameter λ listed in Table II does not play the role attributed to it in the algorithm. Please either remove the exponent from (15) or re-derive the weight update (and check the corresponding H update in Eq. (19)) for the w_k^λ form.","section":"Section III-D, Eqs. (15), (20a)-(20d)"},{"comment":"The empirical comparison is weakened by test-set model selection. Section IV-A says that parameters are set by grid search over the ranges in Table II, but no validation split or independent model-selection procedure is described; the means reported in Tables IV and V are computed on the same datasets used to select the parameters. This is especially problematic for HSS-MVFC, which has four free parameters (m, λ, η, r), more than most baselines, so its rank advantage may partly reflect overfitting to the test sets. In addition, the paper's own Holm post-hoc tests (Tables VIII-IX) fail to reject equality with MV-Co-FCM, Co-FKM, TW-K-means, and MVKKM on both RI and NMI, so the support for 'better than many related methods' is narrower than the abstract claims. Please report a proper model-selection protocol (e.g., inner cross-validation) and the selected parameter values per dataset, and temper the conclusions accordingly.","section":"Section IV-A/C, Tables II, IV, V"},{"comment":"The algorithm description is not reproducible as written. Step 9 of Table I says 'Until (7) reaches a minimum', but Eq. (7) is the Co-FKM objective from Section II-B, not the HSS-MVFC objective (15); the stopping criterion should refer to (15). The initialization of P_k and H is never specified, no stopping tolerance is given, and the convergence evidence in Fig. 3 is empirical only, with no convergence proof for the alternating scheme. These omissions matter because the reported numbers in Tables IV and V depend on initializations and stopping rules.","section":"Section III-E, Table I"}],"minor_comments":[{"comment":"The displayed objective (15) is typeset incorrectly, with superscripts, subscripts, and constraints scrambled; please provide a clean display of the objective and all constraints.","section":"Section III-C, Eq. (15)"},{"comment":"The grid for λ in HSS-MVFC is printed as '{2-3,2-2,...,2,29,210,211,...,214}'; it should use superscripts (e.g., 2^{-3}, ..., 2^9, 2^{10}) to be unambiguous.","section":"Table II"},{"comment":"The ranking of MVSpec is printed as '61667' and should be 6.1667.","section":"Table VII"},{"comment":"Dataset names contain typos: 'Dematology' should be 'Dermatology' and 'Retuters' should be 'Reuters'.","section":"Tables IV and V"},{"comment":"The figure showing convergence curves is labeled 'Fig.4' in the caption but is referenced in the text as Fig. 3; the numbering should be made consistent.","section":"Section IV-D, Fig. 4"},{"comment":"Reference [30] is incomplete (missing venue, volume, pages), and several other references have inconsistent formatting; please unify the bibliography style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The weight-update mismatch between Eq. (15) and Eqs. (20a)-(20d) is central and should be addressed head-on; the empirical protocol also needs a validation-based model-selection procedure. I see no indication of bad faith, and the contribution is potentially publishable after a careful revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nHere is my read on Deng et al. The genuinely new thing is coupling a shared hidden space learned by NMF with fuzzy c-means and maximum-entropy view weighting in one alternating objective. Earlier NMF-based multi-view clustering typically builds the latent space first and clusters after; this paper instead alternates so the partition feeds back into the hidden-space learning. That is a sensible, incremental idea, and the paper does more than most in the subfield: full update rules, complexity, six datasets, eleven baselines, Friedman plus Holm post-hocs.\n\nWhere it gets soft: the weight update (20d) does not match the objective as written. If the reconstruction term in (15) is w_k^λ ||X_k−P_kH||^2, the Lagrangian minimizer is not a simple softmax over D_k unless λ=1. Either the objective or the update needs correcting. That is not a minor notation nit; it means the algorithm may be solving a different problem than claimed, at least for the grid values of λ they search.\n\nSecond, parameter selection. Table II gives grid ranges for every method, and the reported numbers are means over 10 runs at chosen settings, but there is no mention of a validation split or hold-out selection. Everything is tuned on the same datasets used for the final comparison. HSS-MVFC has four free parameters (m, λ, η, r), more than most baselines, so its rank advantage in Tables VI–VII is plausibly inflated. The Holm tests tell the same story: HSS-MVFC is significantly better than roughly half the baselines, not all, which is a far more modest claim than \"mean highest on all datasets.\"\n\nThe paper also ships no code and does not report the chosen parameter values, so an independent check is not possible. Some equation-numbering and table typos make the text harder to trust than it should be.\n\nThe core idea is worth engaging with. I would not desk-reject this: a serious referee can ask for the weight-update derivation to be fixed, a proper validation protocol, and code or at least selected parameter values. If those are addressed, it is a reasonable niche contribution. For my own work, I would not cite it as-is.","headline":"A plausible incremental multi-view clustering method whose empirical claim is weakened by test-set parameter selection and a weight-update mismatch with the stated objective.","tokens_in":17130,"tokens_out":8756,"would_cite":false,"duration_ms":80830,"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":"Cluster in a shared hidden space learned jointly with the fuzzy partition, and multi-view clustering improves on ten existing methods across six datasets.","keywords":["multi-view clustering","fuzzy c-means","non-negative matrix factorization","shared hidden space","maximum entropy","adaptive view weighting","alternative optimization"],"falsifier":"Run HSS-MVFC on a small multi-view dataset and record the objective value (15) after every iteration; if the weight formula is the exact minimizer, the sequence must be non-increasing. Equivalently, fix $\\mathbf{U}$, $\\mathbf{V}$, $\\mathbf{P}_k$, and $\\mathbf{H}$, solve the $\\mathbf{w}$-subproblem by numerical optimization for a case with $\\lambda \\neq 1$, and compare with Eq. (20d): any mismatch means the algorithm optimizes a different objective than claimed.","tokens_in":16267,"feed_emoji":"🧩","tokens_out":6465,"duration_ms":57489,"temperature":0.7,"pith_summary":"The paper proposes a multi-view clustering method, HSS-MVFC, that couples fuzzy c-means (FCM) with non-negative matrix factorization (NMF) so that clustering happens in a hidden space shared by all views. The authors claim that alternately updating the shared hidden space and the fuzzy partition, while assigning each view an adaptive weight through a maximum-entropy criterion, exploits complementary information between views better than clustering views separately or fusing partitions afterward. On six real-world datasets, they report the highest mean Rand Index and Normalized Mutual Information among ten baseline algorithms, with Friedman and Holm tests indicating statistically significant differences for several baselines. If correct, the method shows that representation learning and fuzzy partitioning can be optimized in one objective rather than as separate stages.","feed_headline":"Shared hidden space beats 10 baselines in multi-view clustering","feed_subtitle":"Fuzzy c-means runs inside a shared hidden space while view weights adapt by maximum entropy.","key_machinery":"The load-bearing object is the shared hidden space $\\mathbf{H}$ obtained by factorizing every view as $\\mathbf{X}_k \\approx \\mathbf{P}_k \\mathbf{H}$, so that the same low-dimensional coefficient matrix represents all views and clustering operates on this common space. The updates for $\\mathbf{U}$ and $\\mathbf{V}$ are the standard FCM ones applied inside that space; $\\mathbf{P}_k$ uses the multiplicative NMF update; $\\mathbf{H}$ is updated by a gradient step with a step size chosen to keep the iterate nonnegative; and the view weights take the closed form $w_k \\propto \\exp(-D_k/\\eta)$, where $D_k = \\lVert \\mathbf{X}_k - \\mathbf{P}_k \\mathbf{H} \\rVert_F^2$, so weights fall exponentially with reconstruction error. A maximum-entropy term with parameter $\\eta$ controls how evenly the views are weighted, interpolating between equal weights ($\\eta \\to \\infty$) and winner-take-most ($\\eta \\to 0$).","core_discovery":"The central claim is that a single objective can jointly learn a shared hidden representation of multi-view data and a fuzzy partition of it, and that doing so improves clustering. The objective minimizes FCM distortion in the hidden space, a weighted NMF reconstruction error that forces all views to share one coefficient matrix $\\mathbf{H}$, and a negative Shannon-entropy term that adaptively balances view weights. The authors solve it by cycling through five subproblems: updating memberships, cluster centers, per-view basis matrices, the shared hidden matrix, and view weights, so that the hidden space and the partition are refined against each other. Their experiments report mean RI of 0.8322 and mean NMI of 0.5308 across the six datasets, the highest in both tables.","pith_inferences":["Because the stated objective raises each view weight to the power $\\lambda$ in the reconstruction term while the derived update treats the term as linear in $w_k$, the implemented algorithm may be minimizing a slightly different objective when $\\lambda \\neq 1$; a numerical comparison against direct optimization of the weight subproblem would settle this.","The shared hidden matrix $\\mathbf{H}$ could let the method handle partially missing views: once $\\mathbf{H}$ is learned from complete views, the per-view basis matrices can reconstruct or impute absent views, although the paper does not test this.","The entropy parameter $\\eta$ could double as a diagnostic: shrinking $\\eta$ reveals which views carry the most clustering signal, since their weights dominate.","The same joint objective idea could be transferred to other prototype-based learners, such as possibilistic or noise-robust fuzzy clustering, by swapping the FCM distortion term while keeping the shared hidden space and entropy weighting."],"forward_implications":["Multi-view clustering can be treated as one coupled optimization problem, so improvements in the shared representation directly improve the partition and vice versa.","View weights do not need to be specified in advance; they emerge from each view's reconstruction quality under an entropy-controlled distribution.","A single parameter $\\eta$ lets practitioners trade off between democratic weighting and dominance by the most informative view.","The alternating scheme gives a concrete recipe for adding a shared latent space to other prototype-based clustering models.","The reported mean RI and NMI rankings suggest the benefit is consistent across datasets, not limited to one domain."],"supporting_citations":[{"why":"Supplies the classical fuzzy c-means model whose objective is the base framework for the proposed method.","marker":"[9, 10]"},{"why":"Introduces joint non-negative matrix factorization for learning a shared hidden space across views, the idea HSS-MVFC extends by coupling it to clustering.","marker":"[36]"},{"why":"The Co-FKM baseline that also builds multi-view fuzzy clustering on FCM, used for comparison and motivation.","marker":"[27]"},{"why":"Provides the maximum entropy clustering mechanism that motivates the adaptive view-weighting term.","marker":"[11]"},{"why":"Supplies the multiplicative update rule used for the per-view basis matrices in the NMF subproblem.","marker":"[31]"},{"why":"Supports the claim that maximizing entropy drives view weights toward equality, which justifies the regularizer's behavior.","marker":"[30]"},{"why":"A collaborative fuzzy clustering baseline that also assigns weighted views, used as a comparison and as a related weighted-view formulation.","marker":"[6]"}],"fun_headline_variants":["Shared hidden space improves multi-view fuzzy clustering","Multi-view fuzzy clustering with learned shared space","Alternative learning of hidden space and partition wins","Fuzzy clustering boosted by shared hidden space","Outperforms baselines via shared hidden space"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the assumption that the simple update formula for each view's weight is the exact solution of the weight subproblem, but the derivation is only valid if the reconstruction error appears linearly in the weight, while the objective as written raises the weight to the power $\\lambda$.","fun_headline_variants_meta":{"raw":{"variants":["Shared hidden space improves multi-view fuzzy clustering","Multi-view fuzzy clustering with learned shared space","Alternative learning of hidden space and partition wins","Fuzzy clustering boosted by shared hidden space","Outperforms baselines via shared hidden space"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001029,"raw_usage":{"total_tokens":4283,"prompt_tokens":843,"completion_tokens":3440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":3387}},"tokens_in":459,"tokens_out":3440,"duration_ms":24794,"temperature":1.0,"reasoning_tokens":3387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:48:37.639026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HSS-MVFC on a small multi-view dataset and record the objective value (15) after every iteration; if the weight formula is the exact minimizer, the sequence must be non-increasing. Equivalently, fix $\\mathbf{U}$, $\\mathbf{V}$, $\\mathbf{P}_k$, and $\\mathbf{H}$, solve the $\\mathbf{w}$-subproblem by numerical optimization for a case with $\\lambda \\neq 1$, and compare with Eq. (20d): any mismatch means the algorithm optimizes a different objective than claimed.","supporting_citations":[{"cited_title":"Multiview Partitioning via Tensor Methods","cited_arxiv_id":null,"evidence_quote":"Introduces joint non-negative matrix factorization for learning a shared hidden space across views, the idea HSS-MVFC extends by coupling it to clustering."},{"cited_title":"Convergence of t he Single-Pass and Online Fuzzy C-Means Algorithms","cited_arxiv_id":null,"evidence_quote":"The Co-FKM baseline that also builds multi-view fuzzy clustering on FCM, used for comparison and motivation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the maximum entropy clustering mechanism that motivates the adaptive view-weighting term."},{"cited_title":"The possibilistic C -means algorithm: insights and recommendations","cited_arxiv_id":null,"evidence_quote":"Supplies the multiplicative update rule used for the per-view basis matrices in the NMF subproblem."},{"cited_title":"A possibilistic approach to clu stering","cited_arxiv_id":null,"evidence_quote":"Supports the claim that maximizing entropy drives view weights toward equality, which justifies the regularizer's behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A collaborative fuzzy clustering baseline that also assigns weighted views, used as a comparison and as a related weighted-view formulation."}],"review_version":1}