{"id":"9446a4d9-3c9c-4c43-955d-3324d6c5cdb9","arxiv_id":"2412.01063","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MuSiCNet combines multi-scale attention with Lomb-Scargle periodograms and dynamic time warping to represent irregularly sampled multivariate time series, reporting strong results across three tasks.","lead":"A new machine-learning architecture, MuSiCNet, analyzes irregularly sampled time series by viewing them at multiple time scales from coarse to fine. The authors report results competitive with state-of-the-art on classification, interpolation, and forecasting benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 reports a 4–5× interpolation MSE advantage over mTAND-Full; this is implausibly large and the protocol is underspecified, so the three-task 'consistent SOTA' claim rests on an unverified comparison.","rationale":"The reader's weakest_assumption is LSP-DTW, but the Table 5 ablation shows removing the correlation matrix only drops P12 AUROC from 86.1 to 85.4, so that assumption is not the most load-bearing part of the central claim. The reader's rationale does flag the anomalously large interpolation numbers, which is exactly where I believe the central claim is least secure. I therefore partially agree. The concern is not an accusation; it is an alternative explanation that the current manuscript cannot rule out. A shared-protocol rerun of Table 2 would settle it. I keep the reader's CONDITIONAL verdict because code release, significance testing, and corrected tables are also needed before acceptance, but the decisive check is the interpolation protocol.","tokens_in":18041,"tokens_out":10147,"duration_ms":92656,"concrete_test":"Recompute Table 2 under one shared protocol: fix a single random seed and, for each test instance, randomly mask the stated percentage (50–90%) of observed values once; feed every model (RNN-VAE, L-ODE-RNN, L-ODE-ODE, mTAND-Full, MuSiCNet) the identical remaining observed points with identical normalization, and score all models on the same held-out masked values plus all original time points. If MuSiCNet still achieves MSE ≈ 0.9 while mTAND-Full is ≈ 4.1, the result stands; if the gap collapses to the 10–30% range seen elsewhere, the reported interpolation advantage is a protocol artifact and the three-task claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (abstract; Section 4) is that MuSiCNet is competitive with SOTA consistently across classification, interpolation, and forecasting. The interpolation pillar is Table 2: MuSiCNet reports MSE 0.918–0.992 while the best baseline, mTAND-Full, reports 4.018–4.798 across 50–90% observed levels. A 4–5× improvement is far outside the margins in Tables 1 and 3 and most plausibly indicates a protocol mismatch. Section 4.2 states models reconstruct values for the entire dataset from a selected subset of observed points, but MuSiCNet's interpolation loss (Eq. 10) is only the masked reconstruction at the finest scale, and the paper does not specify how the 'observed %' masks are chosen for MuSiCNet versus baselines, how test-time evaluation points are selected, or whether normalization constants are identical. If baselines are scored on all original time points while MuSiCNet is scored on masked points, Table 2 does not measure the same task. Because the 'consistent SOTA' phrasing also conflicts with Table 1 P19 AUPRC and Table 3 MIMIC/Physionet12, interpolation carries a disproportionate load. No code or baseline re-run protocol is supplied, so the comparison cannot currently be verified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces MuSiCNet, a multi-scale and multi-correlation attention framework for irregularly sampled multivariate time series. The method first converts each instance into a hierarchy of relatively regular coarse-to-fine series via masked average pooling, then encodes each scale with an attention module whose inter-series attention scores are reweighted by a frequency-domain correlation matrix computed with Lomb-Scargle periodograms and dynamic time warping (LSP-DTW). Adjacent scales are tied together by a reconstruction-adjustment loss and a contrastive loss. The framework is evaluated on classification (P19, P12, PAM), interpolation (PhysioNet), and forecasting (USHCN, MIMIC-III, Physionet12), and the authors claim that it is consistently competitive with state-of-the-art across all three tasks.","tokens_in":18328,"tokens_out":8313,"duration_ms":74160,"significance":"The core idea, treating irregularity as relative and learning from coarse to fine through a hierarchy of regularized views of the same instance, is interesting and reasonably well motivated. The paper combines this with an original LSP-DTW correlation estimator and a broad comparison across three tasks, and the ablations isolate the main components. If the reported results can be reproduced under a clearly specified, shared evaluation protocol, the framework could serve as a useful general backbone for ISMTS analysis. However, no code or precise data-processing specification is provided, and the strongest quantitative evidence, the interpolation results, is currently not verifiable. The paper would be strengthened by a reproducible protocol and a more measured statement of the cross-task claim.","major_comments":[{"comment":"The interpolation table carries a disproportionate load in the paper's central claim of consistent state-of-the-art performance, but the reported 4-5x MSE advantage over mTAND-Full is difficult to accept without a precise protocol. The text says models reconstruct values for the entire dataset from a selected subset of observed points, while MuSiCNet's interpolation loss (Eq. 10 in Appendix C) is only the masked reconstruction at the finest scale. The paper does not specify how the observed-percentage masks are generated and applied to MuSiCNet versus the baselines, how test-time evaluation points are selected, or whether all methods use identical normalization constants. Please provide the exact protocol, release the mask-generation and preprocessing code, and, if needed, rerun the comparison under identical evaluation conditions.","section":"Section 4.2, Table 2, Appendix C Eq. (10)"},{"comment":"The abstract's claim of being competitive with state-of-the-art consistently across three tasks is not supported on P19. In Table 1, MuSiCNet obtains P19 AUROC 86.8 versus 89.2 for ViTST and P19 AUPRC 45.4 versus 53.1 for ViTST, a substantial gap relative to the reported standard deviations. Because the cross-task consistency claim is one of the paper's main contributions, the claim should be qualified, or additional evidence should be provided that this level of performance is considered competitive.","section":"Abstract; Section 4.1, Table 1"},{"comment":"Algorithm 1 line 6 passes X^(L) to fCorrE for every scale l, which as written would make all scales operate on the finest-scale series and contradict the coarse-to-fine construction described in Section 3.2 and Figure 2(a). Please clarify whether this is a typo and should read X^(l); if the implementation actually uses X^(l), the pseudo-code must be corrected, because the coarse-to-fine hierarchy is the central mechanism of the paper.","section":"Section 3.2; Algorithm 1, line 6"},{"comment":"The ablation conclusions are stronger than the statistics support. Removing the correlation matrix changes P12 AUROC from 86.1 +/- 0.4 to 85.5 +/- 0.3 and AUPRC from 54.1 +/- 2.2 to 53.0 +/- 2.1, which is within one standard deviation; several other component removals show similarly small deltas. The text states that each component is crucial, but Table 4 does not establish that. Please report confidence intervals or significance tests, or soften the claim.","section":"Section 4.4, Table 4"},{"comment":"The claim that LSP-DTW 'accurately identifies correlations between variables' is not directly validated. Figure 3 is visual and Table 5 measures downstream classification accuracy, which is an indirect proxy. Given that the correlation matrices are computed from series with 88-95% missingness, the LSP-DTW output could be sensitive to the observation window rather than to the underlying signal. Please add a synthetic experiment with known inter-variable correlations or report a quantitative correlation-recovery metric to support the frequency-domain correlation claim.","section":"Section 4.5, Figure 3, Table 5"}],"minor_comments":[{"comment":"The PhysioNet interpolation dataset is described in Section 4.2 and D.2 as containing all 8,000 instances, but Table 6 lists 4,000 samples; please correct this inconsistency.","section":"Section 4.2, Table 6, Appendix D.2"},{"comment":"Physionet12 is described as containing 12,000 patients with an 80.4% missing ratio, while Table 6 lists 5,333 samples and an 85.7% missing ratio; the differing numbers should be reconciled. Also, the PhysioNet challenge dataset in D.2 is cited as [Reiss and Stricker, 2012] but should be [Silva et al., 2012].","section":"Section 4.3, Table 6, Appendix D.3"},{"comment":"The sentence stating that L_int 'essentially represents the reconstruction outcome at the finest scale as l^(L)_adj in Eq. (4)' is confusing because Eq. (4) defines the multi-scale data generation, not a loss; the reference should be to the masked reconstruction loss in Eq. (3) or a restated loss.","section":"Appendix C, after Eq. (10)"},{"comment":"The row labels '(4-6) DTW-based methods' are not actually listed in the table; the table jumps from 'Diag' to 'I-GAK, I-DTW, LSP-DTW'. Please label these rows explicitly for readability.","section":"Table 5"},{"comment":"The hyperparameters lambda1, lambda2, and lambda3 are said to be selected from [1e-3, 1e-2, ..., 1e2], but the chosen values for each dataset are not reported; please include the selected values for reproducibility.","section":"Appendix E.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript is in scope for the journal and the core idea is worth pursuing, but the verification gap in Section 4.2 is substantial, and the absence of code or a detailed data-processing protocol makes it hard to judge whether Table 2 measures the same task as the baselines. I would be willing to review a revision that includes a precise evaluation protocol, corrected dataset statistics, and either code or a detailed re-run specification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know about arXiv:2412.01063: the multi-scale coarse-to-fine framework for irregularly sampled multivariate time series is a genuine contribution, but the paper oversells itself. The interpolation results that carry the \"consistent SOTA\" claim are implausible as reported, and the authors do not give enough protocol detail to rule out a comparison mismatch.\n\nThe new ideas are real. Treating irregularity as relative, building a hierarchy of coarser regular grids, and refining representations from coarse to fine is a sensible organizing principle that I don't think has been applied to ISMTS in this way. The LSP-DTW correlation matrix for re-weighting inter-series attention is also a reasonable response to the spurious-correlation problem in sparse data, and the ablation in Table 5 supports it: LSP-DTW beats interpolation-based I-GAK and I-DTW, as well as trivial baselines. The cross-scale contrastive rectification is well-motivated. The framework is genuinely task-general, covering classification, interpolation, and forecasting in one architecture.\n\nThe soft spots are mostly about evidence quality. The biggest is Table 2. A 4-5x MSE improvement over mTAND-Full (0.92 vs 4.02) is outside anything seen in the other two tasks. The paper says models reconstruct values for the entire dataset from a subset of observed points, but MuSiCNet's interpolation loss is masked reconstruction only (Eq. 10). It never specifies whether baselines are scored on all time points while MuSiCNet is scored on masked points, or how the masks are chosen. That alone makes the \"consistently competitive\" claim unverifiable. Second, the P19 classification numbers are below ViTST (86.8 vs 89.2 AUROC) and Raindrop is essentially tied, yet the abstract claims consistent SOTA. That phrasing needs a downgrade. Third, several ablation deltas on P12 (0.6-0.9 AUROC) fall within the reported standard deviations, so the component-level claims are weaker than the text implies. There are also dataset-statistic inconsistencies between the text and Table 6 (e.g., Physionet12 missing ratio 80.4% vs 85.7%, PAM avg observations 600 vs 4,048). These are minor but should be fixed.\n\nNo code is released, which is a real limitation given the protocol questions. The LSP-DTW reliability concern is valid in principle -- with 78-95% missingness, periodograms can be dominated by sampling windows -- but the Table 5 ablation is direct evidence that it helps, so I would treat that as a minor lingering assumption rather than a flaw.\n\nWho this is for: anyone working on irregularly sampled time series, especially in healthcare. The architecture is worth engaging with, but the results need verification before being taken as established. I'd send it to peer review, but with a demand for code, a corrected Table 2 protocol, and significance testing on the ablations.","headline":"Genuinely new multi-scale architecture for irregular time series, but the 'consistent SOTA' claim rests on an interpolation comparison that is unverifiable as reported.","tokens_in":18860,"tokens_out":3243,"would_cite":true,"duration_ms":26506,"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":"MuSiCNet claims that treating an irregularly sampled multivariate time series as a coarse-to-fine stack of relatively regular series, refined by frequency-domain correlation-weighted attention, is enough to match or beat task-specific…","keywords":["irregularly sampled time series","multivariate time series","multi-scale learning","Lomb-Scargle periodogram","dynamic time warping","attention mechanism","interpolation","forecasting"],"falsifier":"On a synthetic irregular series with known inter-variable correlations and missingness at the level of P12, compare the ranking of pairwise correlations from LSP-DTW against ground truth; if the ranking is no better than chance at 88 percent missingness, or if replacing the LSP-DTW matrix with a fixed identity matrix on real datasets does not hurt AUROC, the central correlation mechanism is not load-bearing.","tokens_in":17818,"feed_emoji":"📈","tokens_out":8058,"duration_ms":65373,"temperature":0.7,"pith_summary":"This paper argues that irregular sampling in multivariate time series is relative, not absolute: by imposing lower, uniform sampling rates on an instance, the same data can be viewed as a stack of increasingly regular series, from coarse to fine. The paper's claim is that building representations by refining from coarse to fine, rather than treating the data as a regular grid with missing values, makes one general model competitive with task-specific state-of-the-art on classification, interpolation, and forecasting. If true, this gives practitioners a single imputation-free backbone for irregular multivariate series, with coarse scales supplying broad temporal context and fine scales preserving detail. The claim is carried by a shared encoder-decoder network whose inter-series attention is re-weighted by frequency-domain correlation matrices computed with Lomb-Scargle periodograms and dynamic time warping.","feed_headline":"Coarse-to-fine scales make irregular time series learnable","feed_subtitle":"Matches specialized models across classification, interpolation, and forecasting, without imputation","key_machinery":"The load-bearing object is the LSP-DTW correlation matrix $C_T$, whose entries are $c_{d_i d_j} = \\min_\\pi \\sum_{(m,n)\\in\\pi} \\left(\\mathrm{LSP}(X_{d_i})[m] - \\mathrm{LSP}(X_{d_j})[n]\\right)^2$, i.e. the dynamic time warping distance between the Lomb-Scargle periodograms of two variables. The Lomb-Scargle periodogram converts unevenly spaced observations of each variable into a power spectrum without imputation, and DTW aligns the two spectra so that shared periodic structure shows up as a small distance and thus high correlation. This matrix re-weights inter-series attention inside the CorrNet encoder at every scale. Around it, the coarse-to-fine hierarchy $X^{(l)}_{\\mathrm{mask}} = M^{(l)} \\odot \\mathrm{AvgPooling}_L(X)$ supplies broad temporal context, and two cross-scale losses, reconstruction adjustment and contrastive consistency, transfer that context into the finer representations.","core_discovery":"MuSiCNet's central discovery is that the irregularity of an irregularly sampled multivariate time series is relative: applying average pooling at decreasing window sizes produces a hierarchy of relatively regular series, and this hierarchy is an asset rather than a nuisance. Within each scale, a multi-correlation attention network encodes the series at reference time points using time attention for intra-series information and a correlation matrix, computed by Lomb-Scargle periodogram followed by dynamic time warping (LSP-DTW), to re-weight inter-series attention; across adjacent scales, reconstruction-output adjustment and contrastive learning rectify the finer representation with the coarser one. The paper reports that this framework sets or matches state-of-the-art results on P12, P19, and PAM classification, reduces interpolation MSE on PhysioNet by a large margin (for example 0.918 versus 4.139 at 50% observed points), and stays within the top two on USHCN, MIMIC-III, and Physionet12 forecasting, with ablations showing that each component contributes.","pith_inferences":["If the coarse-to-fine prior is the main driver of the gains, the same hierarchy should transfer to tasks the paper does not test, such as anomaly detection and online event prediction, where long-term context and sparse sensors matter.","A natural testable extension is to make the correlation matrix learnable per scale, using LSP-DTW as an initialization; if the fixed matrix already captures the signal, a learned correction should change little.","For very short series such as Physionet12, which averages about 130 observations, periodogram estimates are high-variance; a stratified ablation by series length could reveal whether robustness there comes from the multi-scale prior or from LSP-DTW.","If the paper is right, then missingness is partly a modeling choice: by coarsening, one can always produce a regular view, which reframes the field's core difficulty from imputation to choosing the right scale at which to match signals."],"forward_implications":["If correct, imputation is unnecessary for representation learning on irregular multivariate series: coarse scales provide regular structure and LSP-DTW provides inter-variable structure directly from raw timestamps.","A single MuSiCNet backbone can serve classification, interpolation, and forecasting, so practitioners no longer need to pick a task-specific irregular-series model.","The coarse-to-fine refinement implies that broad-view temporal information is a usable prior for fine-grained series, and that adding coarse scales should help most on very sparse data.","The ablations imply that all three components (correlation matrix, reconstruction adjustment, contrastive learning) contribute, since removing all three drops P12 AUROC by about 1.9 points.","Because the correlation matrix is computed once per instance from raw data, the method adds no per-scale imputation overhead and can be applied at 78 to 95 percent missingness."],"supporting_citations":[{"why":"Supplies the least-squares periodogram for unevenly spaced data that LSP-DTW is built on.","marker":"[Lomb, 1976]"},{"why":"Gives the statistical form of the Lomb-Scargle periodogram used as the frequency-domain representation.","marker":"[Scargle, 1982]"},{"why":"Explains the Lomb-Scargle periodogram's properties and supports its use without interpolation.","marker":"[VanderPlas, 2018]"},{"why":"Defines dynamic time warping, the alignment used to turn periodograms into pairwise correlations.","marker":"[Berndt and Clifford, 1994]"},{"why":"Supplies the multi-time attention formulation and reference-point materialization that CorrE adapts, and provides the mTAND baseline.","marker":"[Shukla and Marlin, 2021]"},{"why":"Gives the multiresolution signal decomposition rationale for refining representations from coarse to fine.","marker":"[Mallat, 1989]"},{"why":"Provides the Raindrop baseline and the data preprocessing followed for the classification datasets.","marker":"[Zhang et al., 2021b]"},{"why":"Provides the GraFITi baseline and the observation/forecast-window setup used in forecasting experiments.","marker":"[Yalavarthi et al., 2024]"},{"why":"Supplies the PhysioNet data used in the interpolation and forecasting experiments.","marker":"[Silva et al., 2012]"}],"fun_headline_variants":["Irregularity is relative: coarse-to-fine tames irregular time series","Multi-scale hierarchy turns irregular data into a learning asset","Coarse-to-fine attention refines irregular time series without imputation","Multi-scale correlation attention matches specialized models on irregular series","Seeing the big picture first: multi-scale nets for irregular time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frequency-based correlation fingerprints computed from data that is 78 to 95 percent missing are accurate enough to guide attention; if the fingerprints mostly reflect which times happened to be sampled, the re-weighting could hurt instead of help.","fun_headline_variants_meta":{"raw":{"variants":["Irregularity is relative: coarse-to-fine tames irregular time series","Multi-scale hierarchy turns irregular data into a learning asset","Coarse-to-fine attention refines irregular time series without imputation","Multi-scale correlation attention matches specialized models on irregular series","Seeing the big picture first: multi-scale nets for irregular time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001073,"raw_usage":{"total_tokens":4523,"prompt_tokens":1008,"completion_tokens":3515,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":3427}},"tokens_in":624,"tokens_out":3515,"duration_ms":23020,"temperature":1.0,"reasoning_tokens":3427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:42:53.171072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic irregular series with known inter-variable correlations and missingness at the level of P12, compare the ranking of pairwise correlations from LSP-DTW against ground truth; if the ranking is no better than chance at 88 percent missingness, or if replacing the LSP-DTW matrix with a fixed identity matrix on real datasets does not hurt AUROC, the central correlation mechanism is not load-bearing.","supporting_citations":[{"cited_title":"Least-squares frequency analysis of unequally spaced data","cited_arxiv_id":null,"evidence_quote":"Supplies the least-squares periodogram for unevenly spaced data that LSP-DTW is built on."},{"cited_title":"Studies in astronomical time series analysis","cited_arxiv_id":null,"evidence_quote":"Gives the statistical form of the Lomb-Scargle periodogram used as the frequency-domain representation."},{"cited_title":"Understanding the lomb–scargle periodogram","cited_arxiv_id":null,"evidence_quote":"Explains the Lomb-Scargle periodogram's properties and supports its use without interpolation."},{"cited_title":"Using dynamic time warping to find patterns in time series","cited_arxiv_id":null,"evidence_quote":"Defines dynamic time warping, the alignment used to turn periodograms into pairwise correlations."},{"cited_title":"Multi-time attention networks for irregu- larly sampled time series","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-time attention formulation and reference-point materialization that CorrE adapts, and provides the mTAND baseline."},{"cited_title":"A theory for multireso- lution signal decomposition: the wavelet representation","cited_arxiv_id":null,"evidence_quote":"Gives the multiresolution signal decomposition rationale for refining representations from coarse to fine."},{"cited_title":"Grafiti: Graphs for forecasting irregu- larly sampled time series","cited_arxiv_id":null,"evidence_quote":"Provides the GraFITi baseline and the observation/forecast-window setup used in forecasting experiments."},{"cited_title":"Predict- ing in-hospital mortality of icu patients: The phys- ionet/computing in cardiology challenge","cited_arxiv_id":null,"evidence_quote":"Supplies the PhysioNet data used in the interpolation and forecasting experiments."}],"review_version":1}