{"id":"4482d1a3-c77a-41b1-b0d2-74d98900287f","arxiv_id":"1908.06576","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A bicluster-based visual analysis framework extracts local variable-voxel associations and guides interactive exploration of multivariate scientific data.","lead":"This paper presents a visualization framework that automatically finds groups of variables and voxels with similar patterns, called biclusters, to help scientists explore complex multivariate simulations. The value is that it surfaces local relationships, such as hurricane eye or flame layers, that global correlation analyses miss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bicluster coherence is defined on raw, heterogeneously scaled variables: Eq. 1's pScore lets a single high-range variable dominate, so reported local associations may be scaling artifacts rather than multivariate patterns.","rationale":"After reading the full paper, the central claim turns on the pScore definition in Eq. 1. My concern is not that the system is poorly engineered or the case studies fake, but that the coherence predicate is not invariant under the arbitrary units of the input variables. The paper gives no normalization step and uses a single global delta (20 by default) for variables like pressure (hundreds of hPa), mass fraction (order 0.01), and wind speed (tens of m/s). Under these conditions, the pScore test can be satisfied by a single dominant variable, making the other variables effectively decoration. This would undercut Sec. 7.1's reading of rainbands and the general claim in Sec. 8. The reader's weakest_assumption focused on delta/minv sensitivity; I agree that is a symptom, but the more basic issue is that even the chosen delta has no unit-independent meaning. I keep the reader's CONDITIONAL verdict, and my recommendation is UNCHANGED relative to the reader, not because the concern is minor, but because it is currently a testable risk rather than a demonstrated failure. The proposed normalization check would settle whether the concern lands.","tokens_in":15641,"tokens_out":7715,"duration_ms":83941,"concrete_test":"Re-run the hurricane Isabel extraction after standardizing each variable to zero mean/unit variance (or min-max scaling) and choosing an equivalent tolerance (e.g., delta set to 20 normalized units or derived from a reference variable); compare the association-matrix ranking and the eye/rainband groups. If the top-ranked variable sets or the spatial groups D/E/F change materially, the reported local associations depend on arbitrary variable scaling. A cheaper analytical check: for the biclusters D/E/F in Sec. 7.1, compute the per-variable contribution to pScore; if QVAPOR and VEL contributions are below the noise floor (delta = 20) while PRE accounts for nearly all of the constraint, they are incidental rather than jointly coherent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 1 defines bicluster coherence as pScore = |(s_xu - s_xv) - (s_yu - s_yv)| <= delta, applied directly to raw scalar values of variables with different physical units and ranges (e.g., PRE in hPa vs QVAPOR in kg/kg). If one variable has a much larger absolute range than another, the pScore inequality is approximately |s_xu - s_yu| <= delta for that variable, and the small-range variable contributes almost nothing to the coherence test. A bicluster labeled {PRE, QVAPOR, VEL} can therefore be coherent mainly with respect to PRE, while QVAPOR and VEL vary within a tolerance that is enormous relative to their own scales; the claimed local association among three variables is not actually established. The paper never states that variables are normalized (Sec. 4 introduces s_i,j directly), and Sec. 7.1 uses delta = 20 raw units. The rainband interpretation in Sec. 7.1 (PRE nearly constant, QVAPOR increasing, VEL decreasing) may thus be an artifact of variable scaling rather than a discovered multivariate pattern. Because the central claim is that simultaneous clustering discovers local multivariate associations that global methods miss, this scale dependence is a load-bearing risk: interpretable views are not enough if the biclusters themselves can be generated from a single dominant variable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a co-analysis framework for multivariate scientific data based on biclusters, which are subsets of variables and voxels with close scalar-value relationships. The framework generates all biclusters using a depth-first search with a pScore coherence tolerance (Eq. 1), organizes them by variable sets, groups biclusters by a Jaccard-based similarity metric (Eq. 3), and provides four coordinated views: an association matrix, an MDS-based bicluster view, an enhanced parallel coordinate plot, and a spatial volume rendering. The authors claim this enables users to quickly explore variable sets of interest and discover local correlations of scalar values among variables, and they present three qualitative case studies (hurricane Isabel, turbulent combustion, deep water impact), including one qualitative comparison with GSIM.","tokens_in":15940,"tokens_out":7357,"duration_ms":66123,"significance":"If the methodological caveats are addressed, the framework is a useful contribution to multivariate visualization: it makes biclustering accessible for volume data, provides a clear organizational structure for biclusters, and emphasizes the local nature of variable associations, which global correlation methods miss. The paper is also commendable for explicitly discussing parameter sensitivity and for acknowledging that mathematically correlated scalars may not be physically correlated (Sec. 7.4). However, the central effectiveness claim is currently supported only by qualitative case studies, and two load-bearing technical issues—the scale dependence of the pScore measure and the circularity of the local-correlation statistic—need to be resolved before the experimental observations can be taken as evidence for the claimed local multivariate associations.","major_comments":[{"comment":"The pScore is applied to raw scalar values without any normalization or per-variable scaling. Because the Isabel variables have different physical units and ranges (e.g., PRE in hPa vs QVAPOR in kg/kg), a single tolerance delta=20 is effectively a tolerance on the largest-range variable; the inequality can be satisfied even when small-range variables vary by amounts that are large relative to their own scales. Consequently, a bicluster labeled {PRE, QVAPOR, VEL} does not by itself establish that all three variables are locally associated, since the coherence may be entirely due to PRE. The rainband interpretation in Sec. 7.1 (PRE nearly constant, QVAPOR increasing, VEL decreasing) should be supported by per-variable coherence statistics or by normalizing each variable before biclustering and re-running the analysis.","section":"Sec. 4, Eq. (1); Sec. 7.1"},{"comment":"The correlation C(D') is computed on the voxels of the biclusters of D', which are exactly the voxels selected by the pScore condition. Because pScore <= delta forces s_u - s_v to be nearly constant across the bicluster for every pair (u,v) in D', a strong Pearson correlation between u and v is largely implied by the definition of a bicluster. Thus the high C(D') values used to sort variable sets in the association matrix are partly tautological, and the claim in Sec. 7.2 that local correlation is 'better measured' than global correlation is not supported by this measure. Please report C(D') also on all voxels and on hold-out voxels, and quantify the inflation induced by selecting the bicluster voxels.","section":"Sec. 5.1, Eq. (2); Sec. 7.2"},{"comment":"The evaluation is entirely qualitative: there is no user study, no task-based quantitative metrics, and the only external comparison (GSIM, Fig. 7f) is a single side-by-side visual comparison on one dataset. Moreover, Table 1 shows the number of Isabel biclusters varies from 526 to 4924 as delta goes from 30 to 10, and delta is set to '20 by default' with no stability analysis of the resulting variable sets or groups. Because the central claim (Secs. 7.4 and 8) is that the framework effectively guides exploration, the paper should include a quantitative sensitivity analysis of the top-ranked variable sets and groups across delta and minv, and ideally a small task-based user study comparing exploration outcomes with a baseline.","section":"Sec. 7 and Table 1"}],"minor_comments":[{"comment":"The introduction states that two new experiments concern the ionization front instability data (Sec. 6.1) and hurricane Isabel (Sec. 7.1), but the ionization front data is never presented as a case study in Sec. 7; the only occurrence is the association-matrix screenshot in Fig. 4. Please either add the case study or correct the introduction.","section":"Sec. 1 and Sec. 7"},{"comment":"The text says ten variables were used but lists only nine (PRE, PRECIP, QCLOUD, QGRAUP, QICE, QSNOW, QVAPOR, TC, and VEL). Please add the missing variable or correct the count.","section":"Sec. 7.1"},{"comment":"The double vertical bars in Eq. (1) denote an absolute value in context, not a norm; the notation should be made consistent.","section":"Sec. 4, Eq. (1)"},{"comment":"The text 'sigma_du/sigma_dv' should presumably be the product sigma_du * sigma_dv; the slash notation is misleading.","section":"Sec. 5.1, Eq. (2)"},{"comment":"The choices of 'approximately 10 groups' and the cut-off d(A,B) <= 0.99 are stated without justification; these thresholds affect the exploration and should be discussed or made user-controllable.","section":"Sec. 6.2"},{"comment":"The background-filter threshold of 10% of the total voxels and the minimum-voxel threshold of 0.2% are application-specific; a sentence on how sensitive the displayed features are to these values would help.","section":"Sec. 7"},{"comment":"The GSIM comparison is described only visually; a numerical similarity measure between the extracted groups and the GSIM high-correlation regions would make the comparison more informative.","section":"Sec. 7.3"},{"comment":"The introduction refers to experiments in 'Sec. 6', but the experimental case studies appear in Sec. 7; the cross-reference should be corrected.","section":"Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of Visual Informatics and the core idea is worth publishing after substantive revision. The main concerns are the scale dependence of the coherence measure and the circularity of the local correlation statistic; both affect the validity of the reported findings. I also noticed that the introduction promises an ionization front case study that is absent from the results section, suggesting the journal version was not fully updated from the conference paper. I would not recommend rejection, but the evaluation needs to be strengthened before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, clearly written extension of the authors' SciVis 2018 bicluster paper. The new material is real: a fuller description of the Maple-based generation, a local Pearson correlation defined on bicluster voxels, a density-enhanced parallel coordinate, and two new case studies (hurricane Isabel, ionization front). The coordinated views (association matrix, MDS scatter, parallel coordinates, spatial view) are sensible, and the case studies produce visually coherent features that match domain expectations. If all you need is a practical tool for exploring multivariate volumes, this is a reasonable contribution.\n\nThe main soft spot is the bicluster coherence definition. Eq. 1 uses pScore on raw scalar values, and the paper never normalizes variables. With variables like PRE (~1000 hPa) and QVAPOR (~kg/kg), delta=20 means the QVAPOR term in the pScore inequality is effectively zero, so the bicluster constrains PRE (and any other large-range variable) while ignoring QVAPOR entirely. The claim that a bicluster over {PRE, QVAPOR, VEL} captures a local multivariate association is therefore not established; it may be a univariate constraint on PRE plus an unconstrained QVAPOR. The rainband interpretation (constant PRE, increasing QVAPOR, decreasing VEL) may still be true physically, but the framework didn't automatically extract it as a three-variable coherence. This is load-bearing because the central pitch is simultaneous clustering of variables and voxels. It is fixable: normalize each variable (z-score, min-max, or rank) before computing pScore, or use a per-variable tolerance.\n\nOther weaknesses are lighter. The evaluation is entirely qualitative, with no user study, no quantitative ground truth, and only one GSIM comparison on one dataset. The user-chosen parameters delta and minv strongly control the output (Table 1 shows an order-of-magnitude range in bicluster count), and the paper's guidance is vague. No code or data is provided, which limits reproducibility. The correlation measure in Eq. 2 is computed on the same voxels selected by pScore, so it partly reflects the selection criterion; the paper acknowledges only the user-verification caveat.\n\nThe citation pattern is fine, including the self-citation to [12], which is clearly disclosed and honest.\n\nWho this is for: visualization researchers building multivariate analysis tools, and reviewers who want to see a clear system with suggestive, not definitive, evidence. It deserves a serious referee, but I would not accept it as is. I'd ask for variable normalization (or strong justification for raw scales) and a more rigorous evaluation, even just a small user study or quantitative comparison on synthetic data.","headline":"A useful incremental systems paper, but the bicluster coherence measure is scale-dependent and likely lets one variable dominate; needs normalization or a unit-aware design before the central claims hold.","tokens_in":16438,"tokens_out":5180,"would_cite":false,"duration_ms":54262,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that biclusters—variable plus voxel subsets with close scalar-value patterns—can serve as the organizing unit for exploring local associations in multivariate scientific data.","keywords":["multivariate data","bicluster","local association","co-analysis","scientific visualization","parallel coordinates","correlation analysis","exploratory visualization"],"falsifier":"Run the extraction on a synthetic volume with planted regions of known correlated variables and measure whether recovered biclusters match the planted regions as $\\delta$ ranges from 10 to 30. If no single $\\delta$ recovers both a large background region and a small embedded feature without a tenfold change in bicluster count and a loss of the small feature, the central claim loses support.","tokens_in":15466,"feed_emoji":"🔬","tokens_out":6381,"duration_ms":62151,"temperature":0.7,"pith_summary":"Multivariate scientific data usually hide associations that only hold in part of the volume and for a subset of variables. This paper tries to establish that automatically extracted biclusters—subsets of variables together with subsets of voxels whose scalar values follow a close, similar pattern—can serve as the organizing unit for exploring those local associations. The authors argue that clustering variables and voxels simultaneously finds local relationships that global correlation metrics and manual transfer-function design miss, and that organizing biclusters by variable set and grouping them by spatial overlap makes the result interpretable. If the claim is right, a scientist can start from an unexamined simulation, let the framework enumerate meaningful variable-voxel associations, and then verify a shortlist of groups in the spatial and scalar-value views.","feed_headline":"Biclusters expose local variable links global analysis misses","feed_subtitle":"Clustering variables and voxels together reveals where associations actually live in the volume.","key_machinery":"The load-bearing object is the bicluster and its defining pScore. For a $2\\times 2$ submatrix of two voxels and two variables, pScore is the absolute difference between the two value changes; a bicluster is a set of variables and voxels in which every such pScore is at most $\\delta$. The generation step is a depth-first variable-enumeration search that starts from all variable pairs, expands variable sets, and intersects voxel sets, keeping only closed biclusters above a minimum voxel count. Organization and grouping rest on two simpler metrics: Jaccard similarity between bicluster voxel sets for grouping, and a minimum pairwise Pearson correlation over the bicluster voxels for ranking variable sets. These metrics are what make the bicluster output navigable instead of a flat list of thousands of overlapping regions.","core_discovery":"The paper's central claim is that a bicluster $(D',V')$—defined by the requirement that the pScore of every pair of voxels on every pair of variables in the set be at most a tolerance $\\delta$—is the right primitive for local co-analysis of multivariate data. pScore measures the difference between two voxels' value changes across two variables; a small value means the two voxels follow the same scalar-value pattern on those variables. The framework enumerates all closed biclusters through a depth-first expansion of variable sets, intersects voxel sets as variables are added, and discards biclusters with fewer than $\\mathrm{minv}$ voxels. These biclusters are then organized by variable set, and biclusters sharing a variable set are hierarchically grouped by the Jaccard similarity of their voxel sets, so that users explore representative groups rather than thousands of overlapping regions. A new local correlation measure—the minimum absolute Pearson coefficient among variable pairs over the bicluster voxels—ranks variable sets and can disagree with global correlations computed over all voxels. Four coordinated views, including an association matrix, a bicluster similarity scatter plot, an enhanced parallel coordinate, and a spatial volume rendering, link variable-set selection, bicluster grouping, scalar-value coherence, and spatial distribution.","pith_inferences":["Editorial inference: the bicluster output could directly seed multi-dimensional transfer functions, since each bicluster already encodes a variable subset plus a scalar-value combination—the input such transfer functions need.","Editorial inference: because Table 1 shows the bicluster count varying by nearly an order of magnitude with $\\delta$, an adaptive tolerance that scales with local data range would be a natural stress test and likely improvement; the paper's fixed $\\delta$ is a simplifying choice, not a necessity.","Editorial inference: a quantitative benchmark that compares recovered bicluster regions against hand-labeled features (e.g. hurricane eye and rainbands) would measure precision and recall and would settle whether the groups are truly meaningful.","Editorial inference: the framework's logic suggests it could track associations over time by comparing bicluster variable sets across time steps; the paper lists time-varying data as future work."],"forward_implications":["For a data set like hurricane Isabel, the framework can separate the hurricane eye ($\\{PRE,QVAPOR,TC\\}$) from rainbands ($\\{PRE,QVAPOR,VEL\\}$), a distinction that a global correlation over all voxels would blur.","Biclusters sharing a variable set can be grouped by spatial overlap, so a user explores a small number of representative groups instead of thousands of overlapping biclusters.","The local correlation of a variable set, defined as the minimum absolute Pearson coefficient over bicluster voxels, supplies a ranking that can disagree with global correlation rankings and point to regions where variables interact.","The same pipeline extends two-variable association patterns to three or more variables and lets users verify each association in both scalar-value and spatial views.","The framework's usefulness depends on choosing the tolerance $\\delta$ between 10 and 30 (20 by default) and a minimum voxel count; the paper fixes that count at 0.2% of total voxels to capture small features such as the hurricane eye."],"supporting_citations":[{"why":"Supplies the pattern-based bicluster search with completeness and overlap guarantees used to generate all biclusters.","marker":"[31]"},{"why":"Motivates local scalar-value associations between variables and is extended here from pairs to multivariate biclusters.","marker":"[2]"},{"why":"Provides the gradient similarity measure (GSIM) used as a comparison baseline in the deep water impact experiment.","marker":"[6]"},{"why":"Represents global information-theoretic correlation analysis that the paper contrasts with local bicluster-based correlation.","marker":"[16]"},{"why":"Establishes variance minimization biclustering as the method family the paper adopts.","marker":"[30]"},{"why":"Supports the claim that pScore is a rigorous and noise-robust pattern definition among clustering alternatives.","marker":"[32]"},{"why":"Provides the deep water impact simulation data set used in the third experiment.","marker":"[36]"}],"fun_headline_variants":["Biclusters spotlight local variable links global eyes miss","Co-analysis via biclusters reveals hidden local associations","Local variable-voxel patterns surface with bicluster co-analysis","Bicluster framework exposes region-specific variable correlations","Go beyond global: biclusters map local data associations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework stands or falls on whether a single user-chosen tolerance $\\delta$ (plus a minimum voxel count) can make the extracted biclusters line up with genuinely meaningful local phenomena rather than arbitrary value coincidences.","fun_headline_variants_meta":{"raw":{"variants":["Biclusters spotlight local variable links global eyes miss","Co-analysis via biclusters reveals hidden local associations","Local variable-voxel patterns surface with bicluster co-analysis","Bicluster framework exposes region-specific variable correlations","Go beyond global: biclusters map local data associations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000827,"raw_usage":{"total_tokens":3636,"prompt_tokens":989,"completion_tokens":2647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":2568}},"tokens_in":605,"tokens_out":2647,"duration_ms":20991,"temperature":1.0,"reasoning_tokens":2568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:40:12.924836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the extraction on a synthetic volume with planted regions of known correlated variables and measure whether recovered biclusters match the planted regions as $\\delta$ ranges from 10 to 30. If no single $\\delta$ recovers both a large background region and a small embedded feature without a tenfold change in bicluster count and a loss of the small feature, the central claim loses support.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pattern-based bicluster search with completeness and overlap guarantees used to generate all biclusters."},{"cited_title":"Liu, H.-W","cited_arxiv_id":null,"evidence_quote":"Motivates local scalar-value associations between variables and is extended here from pairs to multivariate biclusters."},{"cited_title":"Sauber, H","cited_arxiv_id":null,"evidence_quote":"Provides the gradient similarity measure (GSIM) used as a comparison baseline in the deep water impact experiment."},{"cited_title":"Biswas, S","cited_arxiv_id":null,"evidence_quote":"Represents global information-theoretic correlation analysis that the paper contrasts with local bicluster-based correlation."},{"cited_title":"Oghabian, S","cited_arxiv_id":null,"evidence_quote":"Establishes variance minimization biclustering as the method family the paper adopts."},{"cited_title":"Kriegel, P","cited_arxiv_id":null,"evidence_quote":"Supports the claim that pScore is a rigorous and noise-robust pattern definition among clustering alternatives."},{"cited_title":"Patchett, J","cited_arxiv_id":null,"evidence_quote":"Provides the deep water impact simulation data set used in the third experiment."}],"review_version":1}