{"id":"d6264da1-3923-44d7-b00f-1f44295e4b91","arxiv_id":"2608.04157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MINT stacks per-series recurrence matrices into a tensor and applies CPD to recover interpretable cross-sensor co-clustering patterns in time series.","lead":"This paper proposes stacking self-similarity matrices (Mplots) of many time series into a 3D tensor and decomposing that tensor to reveal patterns shared across sensors. The result is an interpretable way to co-cluster subsequences across large collections of time series without labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cleaning step's preservation of cross-series structure is unverified; an ablation against raw Mplots could falsify MINT's central mechanism.","rationale":"The reader's weakest assumption correctly identifies the Nonnegative Robust PCA cleaning step as the least secure condition for the central claim. The paper itself acknowledges the absence of recovery guarantees in footnote 1, and the synthetic validation is the only direct evidence that the step preserves cross-series co-occurrence while removing baseline self-similarity. I considered whether the lack of baselines or negative controls is more load-bearing, but that concern primarily affects the strength of the contribution rather than the truth of the claim that MINT can co-cluster; if a baseline also passes, MINT might still work, but if the cleaning step discards the signal, the claim is false. Therefore, the cleaning step is the most load-bearing technical assumption. The proposed ablation test is concrete: it directly compares the full pipeline against simplified variants using the paper's own evaluation criteria, and would settle whether the cleaning step is essential and whether it preserves the required structure. The verdict remains CONDITIONAL because the paper's central claim is plausible and the synthetic experiments provide some support, but the missing ablation and the deferred guarantees justify conditions before full acceptance. I agree with the reader's weakest_assumption, and the proposed test would also address the reader's related items about validation and discard policy by making the comparison explicit.","tokens_in":15588,"tokens_out":14727,"duration_ms":140097,"concrete_test":"Run the exact 50-trial protocol of Section 5 on the same four datasets with three variants: (a) the full MINT pipeline; (b) the same pipeline but stacking raw pooled Mplots directly, without Robust PCA or mean-centering; (c) stacking only the sparse components S_i from the Robust PCA step. Compare the P1-P4 pass rates across variants. If variant (b) or (c) matches variant (a) within a pre-specified tolerance (e.g., pass-rate difference <= 0.05), then the cleaning step is not necessary for the claimed co-clustering and the paper's stated rationale is unsupported; if (b) fails while (a) passes, the cleaning step is validated as load-bearing. Also report how many trials were discarded due to NaN Mplots for each variant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MINT co-clusters cross-sensor patterns depends on Section 3 Step 3: per-Mplot Nonnegative Robust PCA into low-rank plus sparse, followed by mean-centering of the low-rank matrices, before stacking into the tensor. If the recovered low-rank component does not concentrate the aligned periodic structure, the tensor will not contain the co-clustering that CPD is supposed to extract. The paper explicitly defers recovery guarantees for the nonnegative-constrained variant and its projected iteration to future work (footnote 1), so the only current support is the synthetic noise-injection study. That study may be unrepresentative: the injected windows are constructed to create block-like low-rank structure in the Mplots, which could survive regardless of whether the cleaning step preserves the specific cross-series alignment claimed in real datasets like OPSD. No ablation or comparison against stacking raw Mplots (or the sparse components) is provided, so the load-bearing role of the cleaning step is untested. If the cleaning step instead filters out the aligned periodic component, the pipeline's output would be dominated by per-sensor baselines or artifacts, and the central claim would fail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes tensorized self-similarity matrices as a new time series data mining primitive: for a collection of N univariate series, it computes an Mplot (pairwise subsequence distance matrix) for each series, applies per-Mplot Nonnegative Robust PCA followed by mean-centering of the low-rank components, stacks these into a 3-tensor, and extracts CPD factors. The authors claim that the resulting factor matrices co-cluster cross-sensor temporal patterns, as demonstrated by an OPSD electricity demand case study and by a synthetic noise-injection robustness study on four real datasets (Taipei MRT, LargeST traffic, Electricity Load Diagrams, CARE wind turbine). In the synthetic study, they plant transformed traces and report that four hand-defined properties P1-P4 are satisfied in 92-100% of 50 trials, rejecting null hypotheses at Bonferroni-corrected significance. The paper includes code and detailed chance-rate bounds in Appendix B.","tokens_in":15781,"tokens_out":10198,"duration_ms":90868,"significance":"If the central claim were established, the paper would introduce a reusable primitive with interpretable decompositions and a clearly specified pipeline, with machine-checkable code and a transparent null model in Appendix B. The idea of treating pooled self-similarity matrices as a tensor to be decomposed is potentially valuable for exploratory analysis of multi-sensor time series, and the authors are appropriately cautious in limiting the claim to 'highly regular datasets containing motifs at regular intervals.' However, the current evidence is not yet sufficient: the pipeline is not compared with any alternative representation or co-clustering method, the role of the RPCA cleaning step is untested by ablation, and the hypothesis tests are framed against an arbitrary 0.7 threshold rather than the analytically derived chance rates. These issues are fixable within the scope of a revision.","major_comments":[{"comment":"The hypothesis tests are specified as H0,i: p_i <= 0.7, but the random-selection model in Appendix B bounds the chance pass rates at at most 0.00584, 0.176, 0.325, and 0.570 for P1-P4, respectively. Rejecting p_i > 0.7 only establishes that the pass rate exceeds an arbitrary threshold, not that MINT outperforms random selection; for P4 a procedure with pass rate 0.71 would reject H0,4 while remaining only marginally above the chance bound of 0.570. The p-values in Table 6.1 should therefore be computed against the chance model (or an exact binomial test on the chance rates), and the selected ranks R for each trial should be reported, since the P4 threshold floor(0.7R) depends on R.","section":"Section 5.3; Appendix B"},{"comment":"The cleaning step, per-Mplot Nonnegative Robust PCA into low-rank plus sparse followed by mean-centering of the low-rank matrices, is the load-bearing mechanism that is supposed to preserve cross-series co-occurrence while discarding per-sensor baseline self-similarity, but it is never isolated. No ablation compares the full pipeline to stacking raw Mplots, stacking only the sparse components, or skipping mean-centering; footnote 1 explicitly defers recovery guarantees for the nonnegative-constrained variant. Because the synthetic windows in Section 5.1 are constructed to produce block-like low-rank Mplot structure, those experiments do not test the preservation of aligned periodic structure in real data. The authors should add ablations (e.g., raw Mplots, sparse-only, no mean-centering, and no RPCA) under the same noise-injection protocol.","section":"Section 3, Step 3; Section 5.1"},{"comment":"The experimental protocol evaluates MINT in isolation and provides no comparison with any baseline method on the same tasks. The paper claims that the tensorized self-similarity matrix 'concentrates cross-series temporal structure' and that this enables analysis that per-series methods cannot, but the only quantitative evidence is a robustness test of MINT against MINT's own planted windows. Baselines such as CPD applied directly to the raw data tensor, CPD on stacked Mplots without cleaning, NMF or PCA on the raw data, RP-GCN, or a Matrix-Profile-based co-clustering method should be run under the same protocol to substantiate the representational claim.","section":"Section 5.1; Section 6"},{"comment":"The OPSD case study is anecdotal: Component 2 is labeled 'Christmas' and Component 3 'seasonal' by visual inspection of the highest |C| loadings, with no quantitative match to holiday dates or to a seasonal decomposition, and several series are imputed with random noise before analysis. The statement that no single-series method finds this separation is not supported by any experiment with a single-series method, and Figure 4.1 alone does not establish that the decomposition separates the two kinds of demand signature. Either add a quantitative evaluation (e.g., correlation of component loadings with known holiday indicators, comparison with per-series Matrix Profile or spectral methods) or recast the section as an illustrative example.","section":"Section 4.1"},{"comment":"The free parameters m, k, s, b, f, and R are fixed per dataset without sensitivity analysis, and the ranks selected by the Kneedle rule are not reported. The threshold in P4 is floor(0.7R), and the random-pass bound in Table B.1 is 0.570 for R=2, so the evidential strength of P4 depends heavily on the rank distribution actually encountered. The authors should report the distribution of R over the 50 trials and provide sensitivity analyses over at least m, s, f, and b to show that the 92-100% pass rates are not an artifact of a single parameter configuration.","section":"Section 5.4; Table 6.1"}],"minor_comments":[{"comment":"The definition of the Mplot refers to AB-Join distance profiles, but the pipeline uses Self-Join Mplots; please state explicitly that the Mplots are computed on z-normalized subsequences and whether the self-join excludes trivial matches.","section":"Definition 3.1"},{"comment":"Trials that yielded NaNs in Mplots were discarded and re-tried on a new seed; please report how many trials were discarded, since systematic discarding could bias the estimated pass rates.","section":"Section 5.4"},{"comment":"Please report the number of NaN-imputed series in the OPSD dataset and the parameters of the random-noise imputation, because random imputation can create spurious structure in the tensor.","section":"Section 4.1"},{"comment":"The table reports p-values only; exact binomial 95% confidence intervals for the proportions would be more informative, especially for the 0.92 estimate on CARE to Compare.","section":"Table 6.1"},{"comment":"The sentence 'We leave formal verification for future work, but we propose that ... constitutes an existence proof' is internally inconsistent; an existence proof cannot be deferred to future work, so please rephrase.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is competently written and the central idea is reasonable, but the current evidence is too self-contained. I recommend major revision rather than rejection because the missing baselines and ablations are well within the scope of the experimental setting already introduced. I would also ask the editor to encourage the authors to report selected ranks and exact binomial intervals, as these are easy to add."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Josh—\n\nRead the MINT paper. The core object is new and it is a real contribution: nobody has stacked per-series Mplots into a tensor and run CPD to co-cluster sensors by shared subsequence patterns. The authors are honest about what is theirs (the primitive, not the decomposition), and the phase-ground-truth figure in Section 1 is persuasive: three phase groups come out as clean blocks in A, and B and C align with the analytically computed offsets. The OPSD case study is a nice existence proof that the factors can separate Christmas demand from seasonal baseline.\n\nThe main soft spot is exactly the one the stress test flags. Step 3 of the pipeline—per-Mplot nonnegative Robust PCA, then mean-centering the low-rank part—is load-bearing. The paper argues mean-centering is needed so CPD doesn't spend a component on per-sensor self-similarity baseline, but there is no ablation showing that the cleaned tensor is actually better than stacking raw Mplots (or the sparse parts) for co-clustering. The synthetic noise-injection study is clever and the hypergeometric null bounds are a legitimate effort, but the injected windows are designed to create block structure that may survive regardless of the cleaning step. As written, the cleaning step's preservation of cross-series structure is untested.\n\nBeyond that, the evaluation has standard fixable gaps: no baseline comparisons, validation thresholds (0.7 null, floor(0.7R) for P4) are hand-picked, and trials that produce NaNs are discarded and re-run on a new seed, which can silently bias the random-selection model. The citation pattern is fine—building on their own Mplot and SCAMP work is legitimate and acknowledged.\n\nThe claim of an 'existence proof' is stronger than the evidence. With one ablation (raw vs. cleaned Mplots) and a baseline or two, this would be a solid paper. As is, it deserves serious refereeing—there is a real primitive here and the experiments are not sloppy—but it needs revision before I'd trust the central mechanism.\n\nTake it? Yes, send to peer review. Expect major revision. For reading group, worth a session.","headline":"The stacked-Mplot tensor is a genuinely new primitive, but the load-bearing Robust PCA cleaning step is never ablated, so the co-clustering claim is not yet fully supported.","tokens_in":16369,"tokens_out":2378,"would_cite":true,"duration_ms":22495,"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 MINT pipeline treats the tensor of stacked Mplots as a mineable object whose CPD components recover shared cross-sensor temporal patterns.","keywords":["time series data mining","recurrence plots","Mplots","self-similarity matrices","tensor decomposition","canonical polyadic decomposition","co-clustering","robust PCA"],"falsifier":"Run the full 50-trial MINT protocol on a dataset of independent random-walk series with no shared motifs: if the top-s sensors and top-s intervals still align in a majority of trials, the co-clustering is a representation artifact. Alternatively, on synthetic data with planted aligned windows, inspect the cleaning output directly: if the planted windows appear in the sparse residual S instead of the low-rank matrix L, the pipeline's central assumption fails.","tokens_in":15350,"feed_emoji":"🧩","tokens_out":6605,"duration_ms":56641,"temperature":0.7,"pith_summary":"This paper introduces a new time series mining primitive: instead of compressing each series to a matrix profile vector, it keeps each series' full pairwise self-similarity matrix (its Mplot) and stacks all of them into a third-order tensor. The MINT pipeline cleans each Mplot with nonnegative robust PCA, mean-centers the low-rank part, and runs canonical polyadic decomposition on the stacked tensor. The claim is that the resulting factor matrices co-cluster cross-sensor patterns: the same components index which sensors, which subsequences, and which time intervals participate in a shared phenomenon. The authors show the pipeline separates Christmas events from seasonal electricity demand across 18 countries, and back it with a 50-trial noise-injection experiment where four co-clustering properties pass in 92-100 percent of trials, far above the chance bounds they derive.","feed_headline":"Stacked Mplot tensors expose cross-sensor co-clusters","feed_subtitle":"A clean-and-decompose pipeline on self-similarity matrices separates Christmas from seasonal demand.","key_machinery":"The load-bearing object is the stacked, cleaned Mplot tensor. An Mplot is the matrix of Euclidean distances between every pair of sliding-window subsequences of one time series, so it records when a series is self-similar to itself at different lags. Each Mplot is pooled, then decomposed by nonnegative robust PCA into a low-rank matrix (periodic self-similarity) and a sparse residual; the low-rank matrices are mean-centered to remove each series' baseline self-similarity level, and stacked into a third-order tensor. CPD then decomposes this tensor into three factor matrices, giving per-component sensor, subsequence, and time loadings. The mean-centering is what keeps CPD from spending a component ranking sensors by overall regularity; the nonnegative low-rank cleaning is what concentrates the time-aligned structure.","core_discovery":"The central discovery is that full self-similarity information, pooled, cleaned, and stacked, concentrates cross-series temporal structure well enough for standard tensor decomposition to extract it. For a dataset of N univariate series of length n with subsequence window m, MINT computes N Mplots, pools them to k×k, splits each into low-rank plus sparse via nonnegative robust PCA, mean-centers the low-rank slices, and stacks them into an N×k×k tensor. CPD factorizes this tensor into a sensor-mode matrix A and two subsequence/time-mode matrices B and C; the authors read each rank-1 component as a shared pattern. In the OPSD case study, two components correspond to Christmas spikes and long-scale seasonal demand, and in the synthetic and injected-noise experiments the components align with planted windows across sensors.","pith_inferences":["The authors leave implicit a boundary condition: the method targets highly regular datasets whose motifs recur at regular intervals; a natural extension would test how component stability degrades as motif periodicity is corrupted.","A testable extension would verify the cleaning step directly: on synthetic data with planted cross-series windows, check that the windows live in the low-rank component rather than the sparse residual or the mean-centered offset.","Because the tensor construction is agnostic to the similarity measure, the same stacking-and-decomposition scheme could be tried with warping-invariant or phase-space self-similarity matrices, which the authors list as future work."],"forward_implications":["Every decomposition component directly indexes the sensors, subsequences, and time intervals participating in a shared pattern, so cross-series structure is recoverable without labels or learned embeddings.","In the four case studies, planted or reference-aligned windows are found in the top loadings in 92-100 percent of 50 random trials, with null hypotheses rejected at Bonferroni-corrected significance.","The OPSD case study shows that short-scale holiday spikes and long-scale seasonal demand separate into distinct components within the same decomposition.","The pipeline's O(N n^2 + N k^3) complexity makes large collections of series analyzable without per-series visual inspection, with robust PCA as the practical bottleneck."],"supporting_citations":[{"why":"Defines the Mplot as a full self-similarity matrix and supplies the primitive this paper tensorizes.","marker":"[22]"},{"why":"Provides the GPU implementation MINT uses to compute Mplots.","marker":"[30]"},{"why":"Supplies the robust PCA formulation MINT adapts with a nonnegativity constraint on the low-rank part.","marker":"[29]"},{"why":"Provides the CPD implementation used to decompose the stacked tensor.","marker":"[11]"},{"why":"Establishes the matrix profile as the basis for subsequence similarity; Mplots extend it to full pairwise distances.","marker":"[28]"},{"why":"Supplies the optimal hard threshold used to estimate the noise level that calibrates the cleaning step.","marker":"[5]"},{"why":"Produces the distance profiles that make up the rows of each Mplot.","marker":"[17]"},{"why":"Provides the Taipei MRT dataset used as one of the four robustness experiments.","marker":"[27]"},{"why":"Provides the LargeST traffic dataset used in the robustness experiments.","marker":"[13]"},{"why":"Provides the Electricity Load Diagrams dataset used in the robustness experiments.","marker":"[25]"}],"fun_headline_variants":["Tensor stacks reveal shared time series motifs","MINT decomposes stacked Mplots for cross-sensor co-clusters","Stacked self-similarity tensors uncover sensor co-clusters","Tensorized recurrence plots expose cross-sensor motifs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that the nonnegative robust PCA cleaning, which the paper notes lacks recovery guarantees, keeps time-aligned periodic structure in the low-rank component of each Mplot while discarding per-sensor baseline self-similarity.","fun_headline_variants_meta":{"raw":{"variants":["Tensor stacks reveal shared time series motifs","MINT decomposes stacked Mplots for cross-sensor co-clusters","Stacked self-similarity tensors uncover sensor co-clusters","Tensorized recurrence plots expose cross-sensor motifs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000489,"raw_usage":{"total_tokens":2373,"prompt_tokens":874,"completion_tokens":1499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1433}},"tokens_in":490,"tokens_out":1499,"duration_ms":9907,"temperature":1.0,"reasoning_tokens":1433,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:41:42.405462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full 50-trial MINT protocol on a dataset of independent random-walk series with no shared motifs: if the top-s sensors and top-s intervals still align in a majority of trials, the co-clustering is a representation artifact. Alternatively, on synthetic data with planted aligned windows, inspect the cleaning output directly: if the planted windows appear in the sparse residual S instead of the low-rank matrix L, the pipeline's central assumption fails.","supporting_citations":[{"cited_title":"Shahcheraghi, R","cited_arxiv_id":null,"evidence_quote":"Defines the Mplot as a full self-similarity matrix and supplies the primitive this paper tensorizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the robust PCA formulation MINT adapts with a nonnegativity constraint on the low-rank part."},{"cited_title":"Kossaifi, Y","cited_arxiv_id":null,"evidence_quote":"Provides the CPD implementation used to decompose the stacked tensor."},{"cited_title":"Mueen, S","cited_arxiv_id":null,"evidence_quote":"Produces the distance profiles that make up the rows of each Mplot."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LargeST traffic dataset used in the robustness experiments."},{"cited_title":"Trindade , ElectricityLoadDiagrams20112014","cited_arxiv_id":null,"evidence_quote":"Provides the Electricity Load Diagrams dataset used in the robustness experiments."}],"review_version":1}