{"id":"4e3f0340-ce7e-4796-8a9a-0197803b5c44","arxiv_id":"2507.19988","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TULCA extends the ULCA dimensionality reduction method to high-order tensors, unifying discriminant analysis and contrastive learning for flexible tensor group comparison.","lead":"This paper introduces TULCA, a tensor decomposition method that lets analysts compare groups of tensors by separating them and highlighting group-specific variance patterns. It also provides interactive visualizations that help analysts steer the method and interpret the results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-mode independent optimization in Eq. 11 assumes group comparisons factor across modes; cross-mode interaction patterns invisible to any single mode-n covariance are not tested, so 'flexible comparison of tensors' is not generally established.","rationale":"The reader's weakest assumption identifies exactly the load-bearing issue: Sec. 3.2.2 optimizes each mode's projection matrix independently on mode-n matricized covariances, then combines the projections via Eq. 4. The paper does not justify why this per-mode factorization is sufficient for tensor-level comparative structure, and it does not test any case where the comparative signal lives in cross-mode interactions. My concrete construction shows a plausible regime where the per-mode covariance matrices are identical across groups even though the tensors are structurally different, so TULCA cannot use group information to steer the projections. This does not invalidate the demonstrated results on the synthetic and real datasets, but it does mean the central claim 'enables flexible comparison of tensors' is broader than what is established. The source code availability, domain-expert case study, and performance evaluation are real supporting evidence, but they do not address this structural limitation. Since the reader already recommended conditional acceptance with clarification on this point, my stress-test does not change the verdict; it sharpens the requested clarification into a specific falsifiable experiment. If the proposed test were run and TULCA still separated the groups, the concern would be refuted; if not, the paper should either narrow its claims or adopt a joint/alternating optimization that respects cross-mode interactions.","tokens_in":21584,"tokens_out":14470,"duration_ms":184604,"concrete_test":"Construct a synthetic third-order tensor with two groups: choose orthonormal A=[a1,a2], B=[b1,b2]; let u_t,v_t be i.i.d. zero-mean with equal variance. For Group 1, set slice X_t = A diag(u_t,v_t) B^T; for Group 2, set X_t = A [[0,u_t],[v_t,0]] B^T. First verify that the mode-2 and mode-3 within-group covariance matrices are equal across groups. Then run TULCA with contrastive weights (w_tg=(1,0), w_bg=(0,1), w_bw=0) and K'_2=K'_3=1. If the core tensors do not separate the groups (or achieve no better separation than a random projection), while an alternating joint optimization of the full multilinear objective (Eq. 4) does separate them, then the independent per-mode optimization in Eq. 11 is the limiting factor and the concern lands. If TULCA still separates the groups, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that solving Eq. 11 independently for each mode, using only mode-n matricized covariance matrices (Eqs. 12-13, mirroring Eqs. 6-7), yields projection matrices whose joint application (Eq. 4) captures the group comparison of interest. This is a genuine factorization assumption: the quantity that matters for comparison, e.g., the between-group contrast of core tensor norms, depends on the joint multilinear mapping, including associations between modes. A group difference can be invisible to every single mode-n covariance: take two groups of 2x2 slices with orthonormal factor matrices A=[a1,a2] and B=[b1,b2], Group 1 having coefficient matrix diag(u_t,v_t) and Group 2 having [[0,u_t],[v_t,0]], with u_t and v_t i.i.d. and E[u_t^2]=E[v_t^2]. Then both groups have identical mode-2 and mode-3 covariance matrices, so Eq. 11 gives no contrastive preference, even though the groups are structurally different. The synthetic and case-study evaluations use patterns that are visible in marginal covariances, so they do not exercise this regime. Because the method's defining novelty is the tensor extension, this untested factorization assumption is load-bearing for the claim that TULCA enables flexible comparison of tensors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TULCA (tensor unified linear comparative analysis), a tensor decomposition method that extends ULCA from matrices to high-order tensors. The method solves a trace-ratio optimization per mode, using weighted within-class, between-class, and background covariance matrices computed from mode-n matricizations, then projects all non-comparison modes via n-mode products to form a core tensor. The authors also propose a CP-based visualization scheme for the core tensor and integrate it into an interactive visual analytics interface. The method is evaluated with synthetic data, two real-world case studies (K computer operational logs and MHEALTH physical-activity data), and runtime benchmarks against Tucker decomposition, with an additional qualitative comparison against ULCA applied to matricized tensors.","tokens_in":21865,"tokens_out":4989,"duration_ms":63202,"significance":"If the claims hold, TULCA fills a real gap: existing tensor decomposition methods do not support contrastive learning, and the ability to blend discriminant and contrastive objectives in one tensor decomposition is useful for visual analytics. The strengths of the paper include a clearly specified optimization, an efficient update strategy that pre-computes covariance matrices, public source code, a sensible core-tensor visualization design, and domain-expert feedback from a realistic supercomputer-log analysis scenario. However, the central claim of \"flexible comparison of tensors\" relies on an untested factorization assumption, and the quantitative evaluation is limited to runtime, with no objective quality metric comparing TULCA against existing tensor discriminant or contrastive methods. These issues are load-bearing for the novelty and scope claims, so the paper needs further work before it can be accepted.","major_comments":[{"comment":"The construction solves Eq. (11) independently for each mode using only the mode-n covariance matrices C^(n)_wil and C^(n)_bwl. This assumes that all group-relevant structure is visible in the marginal mode-n covariances. A counterexample shows the limitation: for two groups of 2x2 slices with orthonormal factor matrices A and B, Group 1 having coefficient matrix diag(u_t,v_t) and Group 2 having [[0,u_t],[v_t,0]] with i.i.d. u_t,v_t and E[u_t^2]=E[v_t^2], the mode-2 and mode-3 covariance matrices are identical across groups even though the multilinear association structure differs. In this regime Eq. (11) gives no contrastive preference, so TULCA cannot separate the groups. The synthetic experiment in Sec. 5 and the case studies in Sec. 6 use patterns that are visible in marginal covariances, so they do not exercise this failure mode. The authors should either add an experiment with a cross-mode interaction pattern of this kind and report the resulting behavior, or explicitly state and discuss this factorization assumption as a limitation of the claim that TULCA enables flexible comparison of tensors.","section":"Sec. 3.2.2, Eqs. (11)-(13)"},{"comment":"The evaluation is almost entirely qualitative. The synthetic experiment in Sec. 5 is assessed visually, and the comparison in Sec. 8 against ULCA is qualitative despite the text in the first sentence of Sec. 8 saying \"We quantitatively compare.\" The only quantitative results are runtimes in Table 2 and Fig. 10. The paper claims to demonstrate TULCA's efficacy, but there is no objective measure of separation quality, variance preservation, or contrastive fidelity, and no quantitative comparison with tensor discriminant analysis (TDA) or a tensor extension of cPCA obtained by matricization. Please add objective metrics for the synthetic and case-study settings (e.g., silhouette scores, classification accuracy, or a between/within variance ratio) and compare TULCA against at least TDA and a matricized-cPCA baseline, or soften the efficacy claims accordingly.","section":"Sec. 5 and Sec. 8"},{"comment":"The visualization pipeline and the parameter sensitivity are not fully validated. The core-tensor visualization in Sec. 4.2 relies on CP decomposition with a default rank R=2, but CP decomposition is subject to scaling and permutation indeterminacy, and the paper does not discuss how sign ambiguity or component ordering affects the scatterplot and bar-chart interpretation. The authors themselves note in Sec. 9 that weight parameter selection requires trial-and-error and can produce spurious visual patterns, but no sensitivity analysis is provided. For a method whose central contribution is interactive steering, a brief sensitivity study showing how stable the core tensor is under small weight changes would substantially strengthen the claims.","section":"Sec. 4.2 and Sec. 9"}],"minor_comments":[{"comment":"Equation (1) says \"identify matrix\"; this should be \"identity matrix.\"","section":"Sec. 3.1.2"},{"comment":"The phrase \"third-core tensor\" should be \"third-order tensor.\"","section":"Sec. 6.1"},{"comment":"There are typographical inconsistencies: \"TULCA-all\" is spelled \"TUCLA-all\" in the text and table header at several points; please unify the spelling.","section":"Sec. 7 and Table 2"},{"comment":"The first sentence says \"We quantitatively compare TULCA and ULCA results,\" but the section title is \"Qualitative Comparison\" and the described evaluation is qualitative. Please correct the wording.","section":"Sec. 8"},{"comment":"The caption contains a typo: \"respecively\" should be \"respectively.\"","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The heavy citation of the authors' own ULCA work is natural because TULCA directly extends ULCA, so I do not treat it as a circularity problem. The more important concern is the scope of the novelty claim: the authors state TULCA is \"the first tensor decomposition method that supports contrastive learning,\" but the paper does not survey or compare against possible tensor contrastive methods beyond TDA references. Please ask the authors to make the factorization limitation explicit and to add a quantitative comparison in the revision. The fit with TVCG is good given the visual analytics interface and the realistic case study, so a major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TULCA is a solid paper and should go to review. It does what it says: it extends the authors' ULCA to tensors with per-mode covariance matrices, adds TcPCA (the tensor analog of contrastive PCA) as a genuinely new piece, and wraps it in a visual interface that makes the core tensor interpretable without dumping an N-mode array on the analyst. The two case studies are the best evidence. The supercomputer log analysis found real operational changes (cooling water temperature shifts, sensor failures) that the domain expert confirmed, and the MHEALTH study shows the same method switching between discriminant, contrastive, and mixed modes. The synthetic experiment cleanly shows each parameter regime behaving as advertised, and pre-computing covariance matrices so weight tweaks cost O(sum K_n^3) instead of a full refactor is a nice practical touch.\n\nThe math is not deep: Eq. 11 solves a trace-ratio problem per mode, exactly parallel to how TDA extends LDA. But the combination with contrastive learning is new for tensors, and the 'first tensor decomposition method supporting contrastive learning' claim is plausible within the literature surveyed. The survey is narrow, though. A referee should ask the authors to check the broader tensor learning literature before standing on 'first.'\n\nThe soft spot worth taking seriously is the per-mode independence assumption. TULCA optimizes each mode's projection separately on mode-n covariance matrices, then multiplies the projections together. If a group difference lives only in the cross-mode structure (identical marginal covariances but different joint structure), the per-mode covariance matrices carry no signal, and TULCA, like TDA before it, is blind to it. The stress-test counterexample is correct. But this is the standard limitation of separable tensor methods, not a flaw peculiar to TULCA, and the paper does not formally overclaim: 'flexible comparison' mainly means switching between discriminant and contrastive schemes. Still, the authors should either test this regime or scope the claim with a sentence.\n\nOther gaps are moderate: no quantitative comparison with TDA or other tensor discriminant methods, and the weight parameters are tuned by trial-and-error until interesting patterns appear. The paper acknowledges both in Sec. 9, including the risk of spurious visual patterns. That honesty counts in its favor. Self-citation is heavy but fair; the code is public and the weights are user inputs, not fitted constants, so circularity is a non-issue.\n\nThis paper is for tensor visual analytics researchers and practitioners working with multivariate time-series data. It deserves a serious referee. I would ask for a cross-mode discussion and a TDA comparison, and I would expect the paper to survive them.","headline":"TULCA is a well-specified tensor extension of ULCA with a real gap-filler in TcPCA and convincing, honest case studies; the untested per-mode factorization assumption is the main caveat but not a dealbreaker.","tokens_in":22401,"tokens_out":7835,"would_cite":true,"duration_ms":86168,"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":"TULCA is a tensor decomposition method that unifies discriminant analysis and contrastive learning, letting analysts compare tensor groups while preserving or suppressing group-specific variance.","keywords":["tensor decomposition","contrastive learning","discriminant analysis","dimensionality reduction","visual analytics","core tensor visualization","supercomputer log analysis","multivariate time-series"],"falsifier":"Construct a third-order tensor whose group labels are separated only by a rank-1 pattern that spans multiple modes, such as a time pattern times an instance pattern times a variable pattern, with no marginal difference visible in any single mode's matricization. If TULCA with discriminant-analysis weights fails to separate the groups in the core tensor while a jointly optimized multilinear projection succeeds, the independent per-mode optimization is the bottleneck.","tokens_in":21406,"feed_emoji":"📊","tokens_out":5500,"duration_ms":54096,"temperature":0.7,"pith_summary":"Comparing groups of multivariate time-series data, such as stable versus unstable supercomputer operation periods, usually requires choosing between two analytic goals: finding factors that separate the groups, or finding factors that are salient in one group relative to another. The paper introduces TULCA, a tensor decomposition method that unifies both goals by extending the matrix method ULCA to high-order tensors. TULCA solves a per-mode trace-ratio optimization with user-tunable weights, so an analyst can emphasize discriminant analysis, contrastive learning, or any mixture of the two. The paper also contributes a visualization scheme that summarizes the resulting core tensor with a 2D scatterplot and bar charts, and a visual analytics interface around it. A sympathetic reader would care because this fills a gap: no previous tensor decomposition method supported contrastive learning, and tensor methods that only separate groups cannot reveal within-group variety.","feed_headline":"First tensor decomposition method supports contrastive learning","feed_subtitle":"TULCA combines discriminant analysis and contrastive learning to separate tensor groups while keeping internal variety.","key_machinery":"The load-bearing object is the per-mode trace-ratio optimization in Eq. 11. For each mode n (n >= 2), TULCA constructs two matrices C_a^(n) and C_b^(n) from mode-n matricizations of the tensor, using group labels along mode-1: C_a^(n) aggregates target within-class covariances and between-class covariances with target weights w_tg^(n) and between-class weights w_bw^(n), while C_b^(n) aggregates background within-class covariances with background weights w_bg^(n). The projection matrix M_n is the maximizer of the trace ratio, solved by iterative eigenvalue decomposition; the projections are then applied through n-mode products to form the core tensor. The weight parameters act as a dial that moves the method continuously between discriminant analysis and contrastive learning. To visualize the N-mode core tensor, the method applies CP decomposition and displays the first two mode-1 factors as a 2D scatterplot and the associated factors of other modes as bar charts.","core_discovery":"The paper's central claim is that TULCA is the first tensor decomposition method that supports contrastive learning, and that by integrating discriminant analysis and contrastive learning it enables flexible comparison of tensors. In the method, for each mode n the projection matrix M_n maximizes the trace-ratio tr(M_n^T C_a^(n) M_n)/tr(M_n^T C_b^(n) M_n), where C_a^(n) is a weighted sum of within-class and between-class covariance matrices of group l and C_b^(n) is a weighted sum of background within-class covariances, with weights chosen by the analyst; applying all projections via n-mode products yields the core tensor. Setting the weights appropriately reproduces TDA, a tensor cPCA (TcPCA), or simultaneous separation and variance preservation, as demonstrated on synthetic data and on supercomputer and mobile-health logs.","pith_inferences":["A natural extension not tested in the paper is joint optimization of all mode projection matrices, such as alternating or coupled optimization, to capture comparative structure that appears only through cross-mode interactions; the independent per-mode optimization is the likeliest place this would matter.","The weight parameters could be learned from analyst demonstrations rather than tuned by hand: the paper cites backward parameter selection for ULCA as a route where an analyst shows a desired scatterplot arrangement and the system infers weights.","For tensors with many groups or modes, the interface's parameter view and per-mode bar charts would need a hierarchical or aggregated design, which the paper itself flags as future work.","The contrastive-learning capability could be tested on other multivariate spatiotemporal domains, such as network traffic or climate sensor arrays, where group-specific salient patterns are the analysis target."],"forward_implications":["Analysts can now compare tensor groups in one decomposition that separates groups and preserves (or suppresses) group-specific variance, a combination no previous tensor decomposition offered.","TULCA reproduces TDA and a tensor extension of cPCA (TcPCA) as special cases of its weight settings, so a single implementation covers both contrastive and discriminant analyses.","Because covariance matrices are precomputed once, updating the weights for interactive steering only requires re-solving the per-mode trace-ratio problem; on the K computer log dataset the update took under a second.","The CP-based core-tensor visualization turns an N-mode core tensor into one 2D scatterplot plus 2(N-1) bar charts, making the results interpretable in a visual analytics workflow.","In the case studies, the method identified cooling-infrastructure changes and sensor failures in supercomputer logs, and differentiated jogging, running, and cycling in mobile-health sensor data."],"supporting_citations":[{"why":"ULCA, the matrix method TULCA extends; supplies the weighted trace-ratio formulation, the iterative eigenvalue solver, and the parameter semantics.","marker":"[20]"},{"why":"TDA, the tensor discriminant analysis baseline whose per-mode covariance construction and projection scheme TULCA generalizes.","marker":"[35]"},{"why":"cPCA, the contrastive learning method whose variance-contrast objective TULCA lifts to tensors as TcPCA.","marker":"[2]"},{"why":"Kolda and Bader's tensor decomposition survey; provides notation, matricization, n-mode products, Tucker and CP decompositions used throughout.","marker":"[33]"},{"why":"Carroll and Chang's CP decomposition, used to factor the core tensor into rank-1 terms for the 2D visualization.","marker":"[11]"},{"why":"Harshman's PARAFAC, the other foundational CP reference used for the same core-tensor factorization.","marker":"[26]"},{"why":"Zou et al.'s contrastive learning via spectral methods, the theoretical basis for the contrastive scheme.","marker":"[64]"},{"why":"LDA, the discriminant analysis formulation that TDA and ULCA build on and that TULCA inherits as a special case.","marker":"[27]"},{"why":"Tucker decomposition, the baseline tensor method compared in experiments and the origin of the core-tensor concept.","marker":"[52]"}],"fun_headline_variants":["TULCA: First tensor decomposition with contrastive learning","Tensor decomposition method merges contrastive and discriminant learning","New tensor decomposition enables flexible comparative analysis","TULCA: Tensor decomposition with contrastive learning for visual analytics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that optimizing each mode's projection independently, on covariance matrices computed from mode-n matricizations defined by mode-1 group labels, captures the group comparison of interest; if the comparative structure lives in interactions among modes, this per-mode optimization could miss it.","fun_headline_variants_meta":{"raw":{"variants":["TULCA: First tensor decomposition with contrastive learning","Tensor decomposition method merges contrastive and discriminant learning","New tensor decomposition enables flexible comparative analysis","TULCA: Tensor decomposition with contrastive learning for visual analytics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2627,"prompt_tokens":899,"completion_tokens":1728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":1676}},"tokens_in":515,"tokens_out":1728,"duration_ms":13239,"temperature":1.0,"reasoning_tokens":1676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:50:33.516653+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a third-order tensor whose group labels are separated only by a rank-1 pattern that spans multiple modes, such as a time pattern times an instance pattern times a variable pattern, with no marginal difference visible in any single mode's matricization. If TULCA with discriminant-analysis weights fails to separate the groups in the core tensor while a jointly optimized multilinear projection succeeds, the independent per-mode optimization is the bottleneck.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TDA, the tensor discriminant analysis baseline whose per-mode covariance construction and projection scheme TULCA generalizes."},{"cited_title":"Gleicher","cited_arxiv_id":null,"evidence_quote":"Harshman's PARAFAC, the other foundational CP reference used for the same core-tensor factorization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Zou et al.'s contrastive learning via spectral methods, the theoretical basis for the contrastive scheme."}],"review_version":1}