{"id":"723667ce-8194-4a46-aafb-1411888fd6c5","arxiv_id":"2411.17354","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DWCL anchors all contrastive pairs to the single highest-silhouette view and uses dual quality and discrepancy weights, reporting gains on eight multi-view clustering benchmarks.","lead":"DWCL is a multi-view clustering method that contrasts every view only against the single best view, selected by silhouette score, and reweights those contrasts by view quality and discrepancy. The authors report consistent accuracy gains over prior deep multi-view clustering methods on eight benchmarks, with the largest absolute gains around 3.5 and 4.4 percentage points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The B-O mechanism's reliance on silhouette coefficient as a view-quality proxy is internally contradicted by Table 7, where SI ranking and per-view accuracy disagree on NUSWIDE.","rationale":"Agree with the reader's CONDITIONAL verdict. The empirical gains are plausible and the ablations support the dual-weighting components, but the core mechanism's premise - that SI is a reliable view-quality signal - is the load-bearing assumption and is directly contradicted by the paper's own NUSWIDE diagnostics. This is not a consensus dispute; it is an internal consistency issue: the weight e^{SI} is supposed to approximate view quality, yet Table 7 shows SI and clustering accuracy rank views differently. A targeted anchor-swap experiment would settle whether the SI choice matters. The theoretical sections do not fix this: Theorem 1's equivalence between weighted InfoNCE and scaled mutual information is an overstatement because InfoNCE is a lower bound, and Theorem 2 establishes only the O(V) complexity reduction by construction. These issues justify the CONDITIONAL verdict: accept only if the anchor-selection premise is validated empirically and the theoretical claims are reworded or corrected.","tokens_in":20701,"tokens_out":7793,"duration_ms":71815,"concrete_test":"Re-run the NUSWIDE experiment with the anchor view fixed to view 4 (lowest SI, second-highest initial ACC) and to each of the other four views, holding all hyperparameters and the dual-weight formula unchanged; if the SI-selected view 5 does not give the best final ACC/NMI, the Best-Other premise fails. For a stronger check, compute Spearman rank correlation between per-view SI and per-view initial ACC across all eight datasets in Table 7-style diagnostics; a non-positive correlation would show the view-quality weight is mis-specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism selects the anchor view by the highest silhouette coefficient (Eq. 8) and weights every cross-view by W_SI = e^{SI} (Eq. 11), so the method's success depends on SI being a faithful unsupervised proxy for view quality. The paper's own Table 7 undermines this premise: on NUSWIDE, view 2 has the second-highest SI (1.0 x 10) but the lowest initial ACC (30.0), whereas view 4 has the lowest SI (0.6 x 10) but the second-highest initial ACC (38.7). Thus the quality weight e^{SI} up-weights a low-accuracy view relative to a higher-accuracy view, and the chosen 'best' view need not be the most informative one. The conclusion concedes this vulnerability. Because all V contrastive pairs are anchored to this single view, an early wrong or unstable SI choice is not a small perturbation: it reweights the entire contrastive objective. The theoretical validation does not resolve this: Theorem 1 treats a weighted InfoNCE loss as equivalent to maximizing mutual information, although InfoNCE is only a lower bound, and Theorem 2 only shows that excluding non-best pairs reduces complexity; it does not show that SI identifies a superior anchor.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DWCL, a multi-view contrastive clustering method. It selects a \"best\" view using the silhouette coefficient (SI), constructs contrastive pairs only between that view and each other view (the Best-Other, or B-O, mechanism), and weights the contrastive loss by the product of a view-quality weight e^{SI} and a view-discrepancy weight derived from normalized mutual information between cluster-label distributions. The authors report consistent improvements over nine baselines across eight multi-view datasets, ablation studies supporting both the B-O mechanism and the dual weighting, an efficiency comparison, and two theorems intended to justify the mechanism theoretically.","tokens_in":20941,"tokens_out":6666,"duration_ms":62205,"significance":"If the empirical results hold, DWCL is a useful and efficient contribution: the B-O mechanism reduces cross-view pairs from O(V^2) to O(V), and the reported 10-run mean±std results consistently favor DWCL over the compared baselines, with ablations showing that both the B-O mechanism and the dual weights contribute. The efficiency comparison is also a positive feature. However, the theoretical validation currently claimed in the paper is not sound: Theorem 1 conflates the InfoNCE lower bound with mutual information equality, and the \"better lower bound\" argument reduces to multiplying a bound by constants; Theorem 2 restates the algorithmic construction rather than proving the superiority of the SI-based anchor choice. Because the authors list theoretical justification as a contribution, these parts must be corrected or substantially reframed before the paper's claims are credible.","major_comments":[{"comment":"The claimed equivalence is not established. The proof cites Oord et al. to assert that minimizing the InfoNCE loss equals maximizing mutual information, but the cited result establishes that InfoNCE is a lower bound on mutual information, not an equality. Moreover, σ is defined in the theorem as the mutual information between high-level representations, while the proof derives the bound on W_CMI using entropies of cluster-label distributions in Eq. (17), so the inequality W_CMI ≥ e^{σ/log N}−1 does not follow from the stated assumptions. The step from Eq. (22) to Eq. (23) therefore multiplies a lower bound by e^{α+β}>1 and calls the result a better lower bound; this is an algebraic restatement, not a proof of improved optimization. In addition, the claim that SI(ĥ_i)∈(0,1) for all points under k-means is not generally true, because silhouette values can be negative, so W_SI∈(1,e) is not guaranteed.","section":"Section 3.4, Theorem 1 (Eqs. 22-23)"},{"comment":"The theorem's conclusion that B-O \"prevents low-quality cross-views from being augmented\" is a direct restatement of the algorithm: by construction, all non-best pairs receive zero weight. The proof establishes only a complexity reduction from O(V^2) to O(V) and a range comparison for the weights; it does not show that selecting the anchor by the highest silhouette coefficient yields a superior or even reliable contrastive objective. Thus the theoretical analysis does not validate the central design choice; it formalizes the mechanism rather than proving its effectiveness.","section":"Section 3.4, Theorem 2"},{"comment":"The method's view-quality measure is the silhouette coefficient, but the paper's own per-view results undermine this premise. On NUSWIDE, View 2 has the second-highest SI (1.0×10) yet the lowest initial clustering accuracy (30.0), while View 4 has among the lowest SI values (0.6×10) yet the second-highest initial accuracy (38.7). Consequently, the weight e^{SI} up-weights a lower-accuracy view relative to a higher-accuracy one, and since the best view is chosen by argmax SI, all V contrastive pairs are anchored to a potentially suboptimal view. The conclusion explicitly concedes that \"the best view may not always be the ideal choice in every scenario.\" The authors should either provide evidence that SI ranking correlates with actual view quality across training iterations and datasets, or replace the criterion with a more robust view-quality measure.","section":"Section 3.2-3.3, Table 7"},{"comment":"The paper claims \"rigorous theoretical justifications\" for both efficiency and effectiveness, but the two theorems as written do not deliver this. Theorem 1 is a tautological scaling of a lower bound, and Theorem 2 is a complexity count plus a definitional zeroing of non-best pairs. The authors should either remove the theoretical claims or substantially rewrite them so that they make falsifiable statements about representation quality, rather than restating the weighting scheme.","section":"Section 3.4 and Section 5"}],"minor_comments":[{"comment":"The abstract in the header reports gains on Caltech6V7 and MSRCv1 of 5.4% and 5.6%, while the full-text abstract reports gains on Caltech5V7 and CIFAR10 of 3.5% and 4.4%; these versions must be reconciled.","section":"Abstract"},{"comment":"The hyper-parameter λ appears both as the multiplier of Σ_v L_v^R in Eq. (14) and inside the definition of L_v^R in Eq. (15), which double-counts the same weight; one of the two appearances should be removed.","section":"Eqs. (14)-(15)"},{"comment":"The text says H(ŷ_v) is the \"cross-entropy\" of the label distribution, but the expression is the Shannon entropy; the terminology should be corrected.","section":"Eq. (13)"},{"comment":"The notation \"W_SIL\" is used in several places instead of the defined \"W_SI\"; the subscripts should be unified throughout.","section":"Table 8 and Section 4.3"},{"comment":"The caption says \"seven baseline methods and BSV,\" but the comparison includes nine named baselines plus BSV; the count should be corrected.","section":"Figure 3 caption"},{"comment":"There is a typo: \"Opitimize\" should be \"Optimize.\"","section":"Algorithm 1, line 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of this paper is defensible and could be publishable after revision. The main liability is the theoretical section: it is presented as a contribution but currently contains a conflation of InfoNCE with mutual information, an unjustified step from weighted losses to scaled mutual information, and a theorem that restates the algorithm. I would ask the authors to either remove these theorems or rewrite them with correct statements and proofs. I would also request a correlation analysis between SI and per-view clustering quality, because the current Table 7 undercuts the paper's stated motivation. Finally, the abstract inconsistency should be fixed before any acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing you need to know: the B-O mechanism is genuinely new and the experiments are mostly convincing. Anchoring all contrastive pairs to the highest-silhouette view, then reweighting by e^SI times SEM's CMI weight, consistently beats pairwise baselines on eight datasets, and the B-O ablation shows it holds. That's a real contribution to deep multi-view clustering.\n\nWhere it's soft: the theory doesn't hold up. Theorem 1 equates the weighted InfoNCE loss with a scaled mutual information term, but InfoNCE is only a lower bound; multiplying a lower bound by e^(alpha+beta) doesn't give a 'better lower bound' unless you show the bound is tight. Theorem 2 mostly shows the complexity reduction, which is true, but the 'safeguarding representation quality' part is not actually proven. The authors state the SI-based best view is chosen by k-means on the current representation, and all pairs are anchored to it. The paper's own Table 7 shows on NUSWIDE the view with the highest SI after pre-training is not the one with the highest ACC, and the conclusion concedes the best view 'may not always be the ideal choice.' That's a real limitation, not a nitpick.\n\nOther soft spots: no code, some baselines are missing on CIFAR10, and the ablation tables (6, 8, 10) don't report variance while Table 5 does. None of that kills the empirical claim, but it means the paper is a bit less reproducible than it advertises.\n\nWho should read it: anyone working on multi-view contrastive clustering. The method is simple enough to implement and the efficiency gain is real. I'd send it to review, but tell the authors to either fix the theoretical section or clearly frame it as a heuristic argument. The empirical core deserves a proper referee.","headline":"A solid empirical multi-view clustering recipe with a defensible new contrastive mechanism, packaged with a theory section that should be ignored.","tokens_in":21520,"tokens_out":1623,"would_cite":true,"duration_ms":15283,"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":"Multi-view clustering anchors all contrastive pairs to one best view, beating pairwise baselines on eight datasets while cutting cost from O(V²) to O(V).","keywords":["multi-view clustering","contrastive learning","representation degeneration","silhouette coefficient","Best-Other mechanism","view quality weight","view discrepancy","deep clustering"],"falsifier":"Find or construct a multi-view dataset where the view with the highest silhouette coefficient has lower true clustering quality (measured by agreement with ground-truth labels) than another view, then run DWCL against the pairwise mechanism: if pairwise achieves higher clustering accuracy or NMI, the central claim that the B-O mechanism anchored by SI reliably improves clustering fails.","tokens_in":1529,"feed_emoji":"🎯","tokens_out":2299,"duration_ms":39096,"temperature":0.7,"pith_summary":"This paper claims that the usual way of building contrastive pairs in multi-view clustering—combining every pair of views—produces many unreliable cross-views and lets low-quality views pull down the shared representation. The authors propose a Best-Other (B-O) mechanism that pairs each view only with the view that has the highest silhouette coefficient, and a dual weighting that downweights cross-views that are both low in quality and high in discrepancy. They argue this mitigates representation degeneration, the phenomenon where contrastive learning makes views too similar and destroys individual-view information. If correct, the method would make multi-view contrastive clustering more accurate, more robust to weak views, and markedly cheaper to run.","feed_headline":"Best-view anchoring lifts multi-view clustering accuracy","feed_subtitle":"Pairing every view with only the highest-silhouette view cuts cost and beats pairwise baselines on eight datasets.","key_machinery":"The central object is the Best-Other cross-view, defined as the pair formed by the view with the highest silhouette coefficient (SI), computed from k-means cluster assignments on each view's representation, and each other view in turn. The silhouette coefficient acts as a view-quality proxy, producing an exponential quality weight $W_{SI}=e^{SI}$; this is multiplied by the view-discrepancy weight $W_{CMI}=e^{CMI}-1$ from prior work, which downweights cross-views whose cluster-label distributions disagree strongly. The resulting dual weight $W_{Dual}=W_{CMI}\\cdot W_{SI}$ is applied to an InfoNCE contrastive loss, and the best view and weights are alternately updated during fine-tuning.","core_discovery":"The central claim is that replacing pairwise cross-view construction with a Best-Other contrastive mechanism, guided by the silhouette coefficient, and weighting each cross-view by the product of a view-quality weight and a view-discrepancy weight, prevents representation degeneration and improves clustering. The paper reports that DWCL outperforms nine prior multi-view contrastive clustering methods and the best-single-view baseline across eight datasets, with absolute accuracy gains of about 3.5% on Caltech5V7 and 4.4% on CIFAR10 over the second-best method, and consistent gains on all views within each dataset. The authors also prove that their dual-weighted loss provides a better lower bound on mutual information than the self-weighted approach and that the B-O mechanism cuts the number of contrastive cross-views from quadratic to linear.","pith_inferences":["A testable extension is to replace the silhouette coefficient with another view-quality estimator (e.g., clustering stability or mutual information with a consensus partition) and check whether performance degrades when the SI-selected view differs from the one chosen by the alternative estimator.","The method implicitly assumes that view quality is a property of the view itself, but quality may be instance-dependent; a view that is globally low-SI could still contain reliable information for specific clusters, which the B-O mechanism would discard.","The abstract in the paper reports 5.4% and 5.6% absolute gains on Caltech6V7 and MSRCv1, while the introduction and experiments report 3.5% and 4.4% on Caltech5V7 and CIFAR10; a reader comparing numbers should verify which datasets and gains the final version intends.","Because the best view is re-selected during training, the anchor can change between iterations, which may cause instability on datasets where two views have near-equal silhouette scores; this is an empirical risk worth probing."],"forward_implications":["If the highest-silhouette view is a reliable quality anchor, multi-view contrastive clustering can be made both faster and more accurate by discarding all non-anchor pairwise cross-views.","The dual weighting should suppress low-quality, high-discrepancy cross-views that otherwise dominate the loss and cause representation degeneration.","Per-view representations should improve uniformly rather than only for the best view, because every view is contrasted against the same high-quality anchor.","Computational cost drops from $O(|V|^2)$ to $O(|V|)$ cross-views, making the method more practical for datasets with many views.","The learned representations transfer to downstream tasks such as linear classification, improving accuracy, recall, and F1 on tested datasets."],"supporting_citations":[{"why":"Defines the Silhouette Coefficient used as the view-quality metric that selects the best view in the B-O mechanism.","marker":"[28]"},{"why":"Supplies the self-weighted contrastive learning framework and the view-discrepancy weight W_CMI that the dual weighting combines with W_SI.","marker":"[41]"},{"why":"Provides the InfoNCE contrastive loss that the dual-weighted objective is built on and the mutual-information interpretation used in the theoretical analysis.","marker":"[26]"},{"why":"Introductes cluster-assignment contrastive learning (CVCL), a key baseline that DWCL compares against and improves on.","marker":"[2]"},{"why":"Represents the DualMVC baseline, which aligns global view features with pseudo-labels and is a state-of-the-art method DWCL is compared with.","marker":"[47]"},{"why":"Represents DCMVC, another recent dual-contrast baseline that DWCL must beat in the empirical comparison.","marker":"[6]"}],"fun_headline_variants":["DWCL pairs only best-other views to boost clustering","Dual-weighting stops representation drift in multi-view clustering","Linear cross-view count: DWCL beats pairwise on 8 sets","Silhouette-guided view pairing lifts clustering accuracy","Cut noisy pairs, gain accuracy: DWCL's dual weights"],"cache_read_input_tokens":23680,"weakest_assumption_plain":"The load-bearing premise is that the silhouette coefficient computed from k-means on the current representation reliably identifies the view that is best to anchor all contrastive pairs; if the highest-SI view is not the most informative one, the entire contrastive mechanism is anchored to a suboptimal view.","fun_headline_variants_meta":{"raw":{"variants":["DWCL pairs only best-other views to boost clustering","Dual-weighting stops representation drift in multi-view clustering","Linear cross-view count: DWCL beats pairwise on 8 sets","Silhouette-guided view pairing lifts clustering accuracy","Cut noisy pairs, gain accuracy: DWCL's dual weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000776,"raw_usage":{"total_tokens":3437,"prompt_tokens":956,"completion_tokens":2481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2409}},"tokens_in":572,"tokens_out":2481,"duration_ms":23573,"temperature":1.0,"reasoning_tokens":2409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:14:55.519236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find or construct a multi-view dataset where the view with the highest silhouette coefficient has lower true clustering quality (measured by agreement with ground-truth labels) than another view, then run DWCL against the pairwise mechanism: if pairwise achieves higher clustering accuracy or NMI, the central claim that the B-O mechanism anchored by SI reliably improves clustering fails.","supporting_citations":[{"cited_title":"Silhouettes: a graphical aid to the interpretation and validation of cluster analysis","cited_arxiv_id":null,"evidence_quote":"Defines the Silhouette Coefficient used as the view-quality metric that selects the best view in the B-O mechanism."},{"cited_title":"Self- weighted contrastive learning among multiple views for mitigating representation degeneration","cited_arxiv_id":null,"evidence_quote":"Supplies the self-weighted contrastive learning framework and the view-discrepancy weight W_CMI that the dual weighting combines with W_SI."},{"cited_title":"Deep multiview clustering by contrasting cluster assignments, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Introductes cluster-assignment contrastive learning (CVCL), a key baseline that DWCL compares against and improves on."},{"cited_title":"Dealmvc: Dual contrastive calibration for multi-view clustering, in: Proceedings of the ACM International Conference on Multimedia, pp","cited_arxiv_id":null,"evidence_quote":"Represents the DualMVC baseline, which aligns global view features with pseudo-labels and is a state-of-the-art method DWCL is compared with."},{"cited_title":"Dual contrast-driven deep multi-view clustering","cited_arxiv_id":null,"evidence_quote":"Represents DCMVC, another recent dual-contrast baseline that DWCL must beat in the empirical comparison."}],"review_version":1}