{"id":"b75d2fe1-142d-4c09-8392-f1878d2dbc89","arxiv_id":"2412.18847","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TPCH improves multi-view hashing clustering by stacking projection matrices and hash codes into tensors and applying an enhanced tensor nuclear norm.","lead":"This paper introduces TPCH, a hashing method for multi-view clustering that learns shared binary codes while enforcing tensor low-rank structure on both the projection matrices and the hash codes. It reports higher clustering accuracy and lower runtime than prior methods on five benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The closed-form Q_p update in Eq. (5) is dimensionally inconsistent and does not solve the stated subproblem, so the paper's claimed optimization of Eq. (2) is not established and the clustering gains are not yet attributable to the method as described.","rationale":"The reader's CONDITIONAL verdict correctly flags dimension inconsistencies in Eqs. (5)-(8) and the missing supplementary theorem. My stress-test sharpens this to a specific, load-bearing algebra error: the closed-form update for Q_p in Eq. (5) is dimensionally inconsistent and does not follow from the stationarity condition of the stated subproblem. This is more fundamental than the missing theorem because it concerns the actual optimization loop in Algorithm 1: if the update is wrong, the algorithm does not minimize Eq. (2), and the empirical results are disconnected from the method's formal description. The code availability is a genuine mitigating factor—if the code implements the correct update, the paper may simply contain transcription errors, and the empirical gains may be real. I therefore do not recommend outright rejection, but conditional acceptance with a mandatory correction of the optimization derivation and a re-verification using the public code. The check I propose directly settles whether the concern lands: monitoring the Eq. (2) objective during runs of the released code. This is a concrete, reproducible test that avoids speculation about intent or hidden assumptions.","tokens_in":13742,"tokens_out":5968,"duration_ms":52698,"concrete_test":"Run the released GitHub code on Cifar-10 with a fixed alpha, logging the value of the objective in Eq. (2) at each iteration; if the objective does not monotonically decrease (within numerical tolerance), then the implemented algorithm is not minimizing the stated objective, and the clustering gains cannot be attributed to TPCH as described. Independently, re-differentiate Eq. (5) to confirm the correct Q_p update is (2 alpha phi(X_p) phi(X_p)^T + mu I)^{-1} (2 alpha phi(X_p) B_p^T + mu A_p - Y_p), not the printed expression.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing weakness is the optimization derivation. In Eq. (5), the subproblem for Q_p is given as alpha * sum_p ||Q_p^T phi(X_p) - B_p||_F^2 + mu/2 * ||Q_p - A_p + Y_p/mu||_F^2, with Q_p in R^{m x l}, phi(X_p) in R^{m x n}, B_p in R^{l x n}, A_p, Y_p in R^{m x l}. The stationarity condition with respect to Q_p yields (2 alpha phi(X_p) phi(X_p)^T + mu I) Q_p = 2 alpha phi(X_p) B_p^T + mu A_p - Y_p. Instead, the paper prints Q_p = [2 alpha B_p phi^T(X_p) + mu(A_p - Y/mu)] / [2 alpha phi(X_p) phi^T(X_p) + mu I], whose numerator is l x m, not m x l, and which implies a right-matrix division. Similarly, Eq. (8) writes B_p = sgn(alpha Q_p phi(X_p) + mu/2 (E_p - J_p/mu)), but the first term must be alpha Q_p^T phi(X_p) for the dimensions to match Eq. (2). These errors mean Algorithm 1, as written, does not minimize the objective in Eq. (2). The paper's central claim that TPCH learns compact hash codes by minimizing this ETNN-regularized objective is therefore not supported. The missing supplementary theorem for the proximal updates of A and E compounds this: the algorithm's behavior is neither verifiable from the paper nor linked to the stated objective. The empirical results in Table 1 are promising and the code is public, but until the optimization is correctly specified (or the code is shown to implement a correct derivation), the reported gains cannot be attributed to the proposed method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TPCH, an anchor- and hash-based multi-view clustering method. It stacks view-specific projection matrices Q_p into a tensor Q and the learned hash matrices B_p into a tensor B, and regularizes both tensors with an 'enhanced tensor nuclear norm' (ETNN) that penalizes a low-rank approximation of the t-SVD core tensor. An ADMM-type algorithm is presented that alternately updates Q_p, B_p, and auxiliary variables A and E; the final hash code is the average of the per-view codes and is clustered by discrete proximal linearized minimization. Experiments on five datasets report large accuracy gains over existing graph- and hash-based methods, together with linear-time complexity and public code.","tokens_in":14117,"tokens_out":8238,"duration_ms":71526,"significance":"If the algorithm as written actually minimized the objective in Eq. (2), the paper would make a useful contribution: the dual tensor low-rank idea on both projection matrices and hash codes is sensible, the linear-time complexity claim is attractive for large-scale multi-view clustering, and the public code is a concrete strength. However, the optimization derivations contain load-bearing errors that prevent the reader from verifying that Algorithm 1 optimizes the stated objective, and the key proximal update is deferred to a supplementary file that is not included. The empirical results are promising but are reported without variance information. The central claim is therefore not currently substantiated, although it appears fixable within the scope of a revision.","major_comments":[{"comment":"The closed-form Q_p update printed immediately after Eq. (5) is not the stationary point of the displayed subproblem. With φ(X_p) ∈ R^{m×n}, Q_p ∈ R^{m×l}, and B_p ∈ R^{l×n}, the first-order condition for minimizing α||Q_p^T φ(X_p) - B_p||_F^2 + (μ/2)||Q_p - A_p + Y_p/μ||_F^2 with respect to Q_p is (2αφ(X_p)φ(X_p)^T + μI)Q_p = 2αφ(X_p)B_p^T + μA_p - Y_p, so the correct update is Q_p = (2αφ(X_p)φ(X_p)^T + μI)^{-1}(2αφ(X_p)B_p^T + μA_p - Y_p). The paper instead prints Q_p = [2αB_pφ^T(X_p) + μ(A_p - Y/μ)] / [2αφ(X_p)φ^T(X_p) + μI], whose numerator is l×m rather than m×l and which divides by a matrix on the right without specifying an inverse. As written, Algorithm 1's step 2 does not solve the Q_p subproblem of Eq. (5), so the algorithm is not shown to minimize Eq. (2). This is a load-bearing error and must be corrected.","section":"Eq. (5)"},{"comment":"The B_p update also contains a transpose error. In Eq. (6), the data term is ||Q_p^T φ(X_p) - B_p||_F^2; expanding this term and dropping the constant binary-norm terms gives a sgn argument of αQ_p^T φ(X_p) + (μ/2)(E_p - J_p/μ), which has dimension l×n and matches B_p. Equation (8) instead prints αQ_pφ(X_p) + (μ/2)(E_p - J_p/μ), which has dimension m×n under the stated dimensions and therefore cannot be an l×n hash matrix. Since this step is the discrete binarization of the hash codes, it is central to the method. Please correct the transpose and present the derivation explicitly.","section":"Eq. (8)"},{"comment":"The enhanced tensor nuclear norm is not sufficiently specified. Equation (3) does not define the low-rank approximation S̄ of the core tensor, the transformation pair B and B^{-1}, or the exact tensor dimensions (the text 'D ×D3' appears to describe a 2-D object rather than a 3-D tensor). More importantly, Update-3 states that the proximal step of A 'can produce a closed-form solution via Theorem 1 provided in supplementary materials,' and Update-4 is dismissed as similar, but the preprint contains no supplementary file and Eq. (9), which Algorithm 1 references for updating A, is missing from the manuscript. Consequently, Steps 4 and 5 of Algorithm 1 are not reproducible from the paper, and the claim that the ETNN-regularized objective is actually being optimized is unverified.","section":"Update-3/Update-4"},{"comment":"The main empirical claim rests on unreplicated point estimates. Table 1 reports one ACC/NMI/Purity value per method per dataset with no standard deviations, no number of trials, and no significance tests. Given the exceptionally large reported margins on Cifar-10 (ACC 0.70 vs next-best 0.29, NMI 0.67 vs next-best 0.13), the absence of variance information makes it impossible to assess whether the improvement is stable or an artifact of a single run or of test-set hyperparameter tuning. Please report means and standard deviations over multiple runs and clarify the hyperparameter selection protocol (e.g., validation-based selection versus selection on the test metric).","section":"Table 1"}],"minor_comments":[{"comment":"The ablation paragraph says the methods without Q and without B are named 'without (w/o) B and without (w/o) B'; the second name should be 'without (w/o) Q', consistent with the legends in Fig. 4.","section":"Ablation Studies"},{"comment":"The caption contains the typo 'Synthtic 3d'; it should read 'Synthetic 3d'.","section":"Fig. 6 caption"},{"comment":"The header of Table 1 has duplicated labels ('Acc Acc', 'NMI NMI') and the column alignment is broken; also, the F-score and ARI results are said to be in the supplementary materials, which are not included in the preprint.","section":"Table 1"},{"comment":"The stopping criterion 'until Satisfy convergence' is vague; please specify a tolerance on the relative change of the objective in Eq. (2) or on the iterates.","section":"Algorithm 1"},{"comment":"The symbol B is used both for the stacked tensor B = Φ([B1; ...; Bv]) and for the binarized matrix constraint B ∈ {-1,1}^{l×n}; the paper should distinguish the tensor, the per-view matrices B_p, and the final averaged hash matrix B̂ defined later.","section":"Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing issues are fixable in a revision: the authors need to correct the transposition/inversion errors in the Q_p and B_p updates, include the missing supplementary theorem for the ETNN proximal step, restore Eq. (9), and add error bars to the experiments. The public code is a clear strength and will help verification. I would not reject the paper, but I also cannot accept it in its current form because the optimization presented in the manuscript does not demonstrably solve the stated objective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: TPCH has a genuinely useful idea and a public codebase, but the paper as written has a load-bearing flaw in the optimization derivation. The Q_p update in Eq. (5) is dimensionally inconsistent with the stated subproblem, and the B_p update in Eq. (8) also mismatches the objective. As printed, Algorithm 1 does not minimize Eq. (2). That's not a nit; it means the reported accuracy gains cannot be attributed to the proposed method from the paper alone.\n\nWhat's actually new: stacking projection matrices into a tensor and applying an enhanced core-tensor nuclear norm to both the projections and the hash codes is a reasonable extension of AC-MVBC. The ETNN idea—low-rank approximation of the t-SVD core tensor with an extra transformation—is plausible, even if the theoretical justification is deferred to a missing supplementary file. The empirical results in Table 1 are strong, especially on Cifar-10, and the public code is a real asset. The ablation study on the two regularizers is sensible.\n\nWhere it's soft: the missing supplementary theorem for the ETNN proximal update is a real gap, because without it the algorithm's behavior is unverifiable. There are also no error bars anywhere, so we can't assess variance. Alpha is tuned on the evaluation metric, which inflates the reported numbers. Those issues are fixable in revision.\n\nThe central problem remains the optimization section. Even if the code implements a correct derivation—and I'd want to run it—the paper itself doesn't give a correct derivation. That's a serious referee point, not a formatting concern.\n\nWho gets value: researchers working on large-scale multi-view hashing clustering. It's a specialized subfield, and the paper is a plausible new baseline. I'd send it to review only if the authors fix the equations and supply the missing theorem. As it stands, it deserves a serious referee but with major revision expected.\n\nMy recommendation: engage with it, but insist on the corrections before accepting the results.","headline":"A useful dual-tensor hashing idea with public code, but the printed optimization equations don't minimize the stated objective, so the paper's accuracy claims aren't yet supported as written.","tokens_in":14695,"tokens_out":4530,"would_cite":false,"duration_ms":37092,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper tries to establish that enforcing an enhanced tensor nuclear norm on both the stacked projection matrices and the stacked hash codes produces more compact and distinguishable binary codes, leading to state-of-the-art clustering…","keywords":["multi-view clustering","binary hashing","tensor nuclear norm","t-SVD","low-rank tensor","bipartite graph","large-scale clustering","Hamming space"],"falsifier":"Run TPCH on a dataset where the cluster-discriminative information is deliberately placed in the high-rank components of the cross-view tensor, for example a synthetic multi-view set where class structure lives in the singular tubes that ETNN truncates; if clustering accuracy does not drop sharply relative to a variant that uses the full core tensor, the low-rank-core assumption is not doing the work claimed, and if it does drop, the assumption is confirmed.","tokens_in":13497,"feed_emoji":"🧩","tokens_out":6692,"duration_ms":53418,"temperature":0.7,"pith_summary":"This paper tries to establish that enforcing an enhanced tensor nuclear norm on both the stacked projection matrices and the stacked hash codes produces binary codes that are more compact, more distinguishable, and more robust to noise, and that this leads to state-of-the-art clustering accuracy on five large-scale multi-view datasets while keeping runtime linear in the number of samples. The core move is to treat the multiple view projections as slices of a third-order tensor and to apply a t-SVD-based low-rank constraint not only to the hash-code tensor but also to the projection tensor, so that views communicate during projection itself. A reader should care because multi-view clustering at large scale is bottlenecked by the quality of low-dimensional codes, and the paper reports a large gain over prior hash-based methods, for example ACC 0.70 versus 0.29 on Cifar-10, with a roughly 7 to 10 times speedup over the strongest competitor.","feed_headline":"Tensor-interacted hashing hits 0.70 ACC on Cifar-10","feed_subtitle":"Stacking view projections and codes into tensors, then low-ranking the core, gives state-of-the-art accuracy in linear time.","key_machinery":"The central object is the enhanced tensor nuclear norm (ETNN), defined on a third-order tensor $T$ as $\\|T\\|_{\\mathrm{etnn}} = \\|\\bar{\\mathcal{S}}\\|_* + \\zeta \\|\\mathcal{U} * \\mathcal{B}^{-1}(\\bar{\\mathcal{S}}) * \\mathcal{V}\\|_*$, where $\\bar{\\mathcal{S}}$ is a low-rank approximation of the core tensor from t-SVD, $\\mathcal{B}$ and $\\mathcal{B}^{-1}$ are linear transforms between core tensors, and $\\|\\cdot\\|_*$ is the nuclear norm of the unfolded tensor. This norm is applied to both the stacked projection tensor $\\mathcal{Q}$ and the stacked hash tensor $\\mathcal{B}$ in the objective (Eq. 2), coupling the per-view projections $Q_p^\\top \\phi(X_p)$ to the per-view binary codes $B_p$; an alternating-direction solver updates $Q_p$, $B_p$ (via a sign function), and the two tensors (via closed-form proximal steps whose theorem is deferred to supplementary). The mechanism is that rank reduction on the core tensor suppresses noise and redundant information in both spaces while the $\\mathcal{B}/\\mathcal{B}^{-1}$ transforms keep the low-rank structure aligned with the t-SVD geometry, yielding hash codes that are more compact and more distinguishable in Hamming space.","core_discovery":"The paper's central claim is that the proposed TPCH method learns more compact and distinguishable hash codes by stacking view-specific projection matrices $\\{Q_p\\}$ into a tensor $\\mathcal{Q}$ and view-specific hash matrices $\\{B_p\\}$ into a tensor $\\mathcal{B}$, and then minimizing the enhanced tensor nuclear norm (ETNN) on both tensors. ETNN differs from the ordinary tensor nuclear norm by first taking a low-rank approximation of the t-SVD core tensor $\\bar{\\mathcal{S}}$ and then applying a pair of transforms $\\mathcal{B}$ and $\\mathcal{B}^{-1}$ that map between the core tensor and its transformed counterpart; the sum of the nuclear norms of these two pieces is what is penalized. The paper argues that this removes noise and redundant information while preserving the higher-order cross-view synergies that the standard tensor nuclear norm leaves unresolved, and that the resulting binary codes, averaged across views and clustered in Hamming space, yield state-of-the-art ACC, NMI, Purity, F-score, and ARI on SUNRGBD, Cifar-10, Caltech101, Caltech256, and 100 leaves, with a time complexity linear in the sample count $n$.","pith_inferences":["A testable extension is to apply the same ETNN dual-tensor trick to other paired-factor models, such as non-negative matrix factorization or dictionary learning over multiple views, where both the dictionary and the codes could be stacked and low-ranked.","The speed advantage hinges on the anchor count $m$ and code length $l$ staying small relative to $n$; on datasets where class structure forces $m$ to grow with $n$, the linear-time claim would erode, so the method's scalability is conditional on anchor sufficiency.","The paper's ablation removes $\\mathcal{Q}$ and $\\mathcal{B}$ separately, but it does not isolate whether the gain comes from core-tensor rank reduction or from the $\\mathcal{B}/\\mathcal{B}^{-1}$ transform; a variant with only the ordinary tensor nuclear norm on both tensors would pinpoint which piece drives the improvement."],"forward_implications":["Multi-view clustering on million-scale datasets becomes practical with substantially better accuracy than prior anchor or hash methods, since TPCH's cost is linear in the sample count.","The dual-tensor constraint yields binary codes that separate clusters more cleanly in Hamming space, so downstream tasks such as nearest-neighbor retrieval or clustering can use cheaper Hamming distances without losing accuracy.","TPCH remains stable under salt-and-pepper noise where the AC-MVBC baseline degrades, suggesting that core-tensor rank reduction is a robustness mechanism usable in other tensor factorization pipelines.","The reported 7 to 10 times CPU-time speedup over the GCAE baseline, if it holds across implementations, removes a practical barrier to large-scale multi-view analysis."],"supporting_citations":[{"why":"Supplies the RBF kernel bipartite-graph construction and the idea of tensor low-rank constraints on stacked binary representations that TPCH builds on.","marker":"(Zhang et al. 2024)"},{"why":"Provides the t-SVD based tensor nuclear norm minimization steps used in the proximal updates for the ETNN tensors.","marker":"(Lu et al. 2019)"},{"why":"AC-MVBC, the main tensor-norm baseline that applies the standard TNN to stacked views; TPCH claims superiority by adding core-tensor low-rank and B/B^{-1} transforms.","marker":"(Zhang et al. 2022)"},{"why":"BMVC, the foundational binary multi-view clustering framework that projects views into a common Hamming space, which TPCH extends.","marker":"(Zhang et al. 2018b)"},{"why":"Supplies the discrete proximal linearized minimization (DPLM) solver used for the final binary clustering step and is the GCAE baseline for runtime comparison.","marker":"(Wang et al. 2023)"}],"fun_headline_variants":["TPCH stacks view projections to sharpen multi-view clustering","Tensor-interacted hashing lifts multi-view clustering accuracy","Cooperative tensor hashing speeds up multi-view clustering","TPCH: tensor co-hashing for state-of-the-art multi-view clustering","Higher-order tensor hashing outperforms on five large datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the low-rank approximation of the t-SVD core tensor, together with the $\\mathcal{B}/\\mathcal{B}^{-1}$ transforms, removes noise and redundancy without discarding the cross-view structure that separates clusters; if rank truncation cuts away class-discriminative signal, the claimed improvement collapses.","fun_headline_variants_meta":{"raw":{"variants":["TPCH stacks view projections to sharpen multi-view clustering","Tensor-interacted hashing lifts multi-view clustering accuracy","Cooperative tensor hashing speeds up multi-view clustering","TPCH: tensor co-hashing for state-of-the-art multi-view clustering","Higher-order tensor hashing outperforms on five large datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000548,"raw_usage":{"total_tokens":2646,"prompt_tokens":1003,"completion_tokens":1643,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":1561}},"tokens_in":619,"tokens_out":1643,"duration_ms":10979,"temperature":1.0,"reasoning_tokens":1561,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:24:57.187219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TPCH on a dataset where the cluster-discriminative information is deliberately placed in the high-rank components of the cross-view tensor, for example a synthetic multi-view set where class structure lives in the singular tubes that ETNN truncates; if clustering accuracy does not drop sharply relative to a variant that uses the full core tensor, the low-rank-core assumption is not doing the work claimed, and if it does drop, the assumption is confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RBF kernel bipartite-graph construction and the idea of tensor low-rank constraints on stacked binary representations that TPCH builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AC-MVBC, the main tensor-norm baseline that applies the standard TNN to stacked views; TPCH claims superiority by adding core-tensor low-rank and B/B^{-1} transforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the discrete proximal linearized minimization (DPLM) solver used for the final binary clustering step and is the GCAE baseline for runtime comparison."}],"review_version":1}