{"id":"f1fc7622-2c99-4b8b-b6d9-9eea49fc70f3","arxiv_id":"2505.07276","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FCPCA performs fuzzy clustering of high-dimensional multivariate time series by projecting each series into cluster-specific common subspaces and minimizing weighted reconstruction error.","lead":"FCPCA is a new fuzzy clustering method for multivariate time series that uses common principal component analysis to reduce dimensions while finding overlapping groups. It could help detect gradual transitions, such as a driver moving from alert to drowsy, in high-dimensional recordings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Varying-length FCPCA does not minimize its stated objective: Eq. (8) drops the (T_i-l) factor implicit in Eq. (7), so Algorithm 1 is not coordinate descent for variable-length series.","rationale":"The strongest claim is that FCPCA is an effective, efficient fuzzy clustering method for high-dimensional MTS with fixed or varying lengths. For that claim, the algorithm's derivation matters, not only its tables. The algebra above shows that the C_s update is inconsistent with the objective whenever lengths differ. This is exactly the paper's advertised variable-length setting. The paper offers no convergence proof, and the unresolved 'see Section??' in Section 4.1.6 plus the one-replicate note already made the reader's verdict conditional. My concern adds a specific, testable reason why the method's internal optimization claim is insecure. I still would not reject: the simulations may remain valid as a heuristic, and the corrected length-weighted update might reproduce or improve the reported RI. That is why the remedy is a comparison of Eq. (7) under both updates, not a blanket dismissal. Agreement with the reader is partial: they flagged the covariance/subspace representation as the weak assumption; I identify a concrete inconsistency in how varying lengths enter that representation.","tokens_in":15983,"tokens_out":11079,"duration_ms":111324,"concrete_test":"Using the authors' GitHub implementation, freeze a membership matrix U from an FCPCA iterate on the Section 4.1.2 Scenario 3 data (lengths 200-600). Compute Eq. (7) twice: once with C_s(l) from Eq. (8), and once with C_s(l) equal to the top-k eigenvectors of S_s^w(l)=sum_i u_{is}^m (T_i-l) hatGamma_i(l). If the second value is strictly smaller, Eq. (8) is not the minimizer of the C-block. Then re-run the full Scenario 3 simulation with S_s^w in Step 9; if RI or convergence behavior changes materially, the published variable-length algorithm is optimizing a different objective than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim is that FCPCA solves Problem (7) by alternating between membership updates and common projection axes. For fixed U, the block of (7) that depends on C_s(l) is sum_i u_{is}^m ||hatX_i(l) - hatX_i(l)C_s(l)C_s(l)^T||^2. Since hatX_i(l)^T hatX_i(l) = (T_i-l) hatGamma_i(l) for the mean-centered block matrix, the minimizer over C_s(l) is the top-k eigenspace of S_s^w(l) = sum_i u_{is}^m (T_i-l) hatGamma_i(l), up to normalization. Algorithm 1, Step 9, instead uses Eq. (8): tildeSigma_s(l) = sum_i u_{is}^m hatGamma_i(l) / sum_i u_{is}^m. The factor (T_i-l) is dropped. With equal series lengths that factor is constant and cancels in the eigenspace; with varying lengths it changes both the projection axes and the subsequent membership updates. Therefore the algorithm is not coordinate descent on the stated objective, no monotone decrease follows, and the convergence check in Algorithm 1 has no formal basis. The length-varying simulations (Section 4.1.2, Scenario 3) are the advertised setting for this feature, yet they are run under an objective different from Eq. (7). This is a concrete algebraic gap, not a tuning issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FCPCA, a fuzzy clustering method for multivariate time series based on common principal component analysis of lagged cross-covariance block matrices. The method iteratively updates fuzzy memberships and cluster-specific projection spaces to minimize a total weighted reconstruction error, with a data-driven criterion for selecting the fuzziness parameter. The authors evaluate FCPCA on simulated VARMA processes, on simulated EEG data, on several benchmark multivariate time series datasets, and on a real EEG driver-drowsiness dataset, comparing against VPCA-based fuzzy clustering, fuzzy C-medoids, and hard CPCA-based clustering.","tokens_in":16306,"tokens_out":4089,"duration_ms":41849,"significance":"If the optimization were consistent, FCPCA would be a useful addition to the fuzzy-clustering literature for multivariate time series: it explicitly addresses variable lengths and high dimensionality, and the paper includes a public R implementation and an interesting real-data application. The comparisons against external baselines (VPCA, FCMD) and against hard CPCA variants are appropriate. However, the central algorithmic claim is undermined by a mismatch between the stated objective and the covariance update used for variable-length series, and the simulation evidence is reported without variability measures and partly contradicts the stated conclusions. These issues currently block acceptance.","major_comments":[{"comment":"For varying-length series, the algorithm does not minimize the objective in Eq. (7). For fixed U, the term involving C_s(l) in Eq. (7) is sum_i u_is^m ||hatX_i(l) - hatX_i(l)C_s(l)C_s(l)^T||^2. Since hatX_i(l)^T hatX_i(l) = (T_i-l) hatGamma_i(l) after mean centering, the minimizer over C_s(l) (for fixed k(l)) is the top-k eigenspace of sum_i u_is^m (T_i-l) hatGamma_i(l), up to normalization. Algorithm 1, Step 9, instead uses Eq. (8), which drops the (T_i-l) factor. With equal lengths this factor is constant and cancels; with unequal lengths it changes both the projection axes and the subsequent membership updates. Consequently, Algorithm 1 is not coordinate descent on Eq. (7), no monotone decrease follows, and the convergence check in Step 15 has no formal basis. Since Scenario 3 in Section 4.1.2 is the advertised variable-length setting, the reported results there are obtained under an objective different from the one stated.","section":"Section 3.1, Eqs. (7)-(8), and Algorithm 1"},{"comment":"The empirical support for the central claim is incomplete. Tables 1 and 2 report only mean Rand indices, with no standard deviations or confidence intervals, so claims of 'outperforms' are not statistically grounded. Additionally, Section 4.1.5 states that only one replicate is used for FCPCA to obtain the RI results, even though Algorithm 1 recommends running multiple random starts; this makes the reported means potentially dependent on a single initialization and weakens the robustness claim. Moreover, the text in Section 4.1.6 says 'FCPCA can detect the fuzzy series in most cases,' but Table 3 reports an average detection rate of 0.00 in several cells (e.g., all three dimension rows for length 400 at threshold 0.6, and the length-200/400 rows at threshold 0.7 for dimension 20). These entries directly contradict the stated claim and the 'fuzzy detection' ability that is a key advertised feature.","section":"Section 4.1.4-4.1.6, Tables 1-3"},{"comment":"The proposed cluster validity index is self-referential. The numerator of Eq. (10) is exactly the objective minimized by the algorithm, and the denominator is built from the same projection matrices P_s = C_s(l)C_s(l)^T that the clustering output produces. Selecting m by minimizing this CVI therefore selects the fuzziness parameter that minimizes the same reconstruction objective, which does not provide independent evidence about cluster validity. The paper should demonstrate on simulated data with known structure that the CVI-based selection of m correlates with an external accuracy measure; the EEG application in Section 5 shows one case where it works, but the claim of 'objectively' selecting m is not established.","section":"Section 3.2, Eq. (10)"}],"minor_comments":[{"comment":"There is a typo in the first sentence: 'Ths study' should be 'This study.'","section":"Section 4.1.1"},{"comment":"The phrase 'improved performance and interoperability' likely should be 'interpretability,' since the paragraph discusses richer clustering insights rather than system interoperability.","section":"Section 1, contribution (4)"},{"comment":"The definition of hatX_i(l) is confusing: the notation X^*_{i,t-l} and X^*_{i,t} suggests columns of a matrix, but the relation to the original X_i is not spelled out. Please clarify the dimensions and the index ranges, since this object is central to the reconstruction error in Eq. (5).","section":"Section 2, Eq. (4)"},{"comment":"Table 3 refers to the 'optimal m' but the procedure for selecting m in the simulations is not defined; the cross-reference 'see Section??' appears to be unresolved. The statement that the most selected m is 'mostly 1.7 or 1.8' would benefit from a table or figure showing the selections.","section":"Section 4.1.6 and Table 3"},{"comment":"The text states that a low fuzziness parameter m=1.1 is used for the real-data comparisons, but Table 6 does not state the value of m used for each method. Please specify the configuration so the results are reproducible.","section":"Section 4.3 and Table 6"},{"comment":"The description of Figure 5 is ambiguous: it says 'the log-transformed CVI values ... The original CVIs and RIs are marked in the figure.' It is unclear whether one plot or two plots are shown and how to distinguish the curves; a legend is needed.","section":"Section 5, Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The algebraic mismatch between Eq. (7) and Eq. (8) is the main technical concern; it is specific and fixable, but it invalidates the convergence claims as stated. The empirical sections also need variability measures and a more careful reconciliation of Table 3 with the text. If the authors can correct the optimization and re-run the variable-length simulations, the paper may become publishable. I do not see grounds for rejection beyond this, as the methodological idea remains plausible and the external comparisons are appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one: the core idea is genuinely new, and there is a load-bearing algebraic gap in the varying-length setting. The new idea is to fuzzify CPCA-based clustering by weighting each cluster's covariance matrix with membership powers, then measuring reconstruction error in the common subspaces. That is a clean extension of the authors' earlier ROBCPCA, and for equal-length series it looks coherent and works reasonably in simulations.\n\nThe problem is with the variable-length claim. Equation (7) defines the objective as sum over i,s of u^m_is || hatX_i(l) - hatX_i(l) C_s C_s^T ||^2. Since hatX_i(l)^T hatX_i(l) = (T_i - l) hatGamma_i(l), the correct weighted per-cluster covariance is sum_i u^m_is (T_i - l) hatGamma_i(l), normalized. Equation (8) drops the (T_i - l) factor. For equal lengths that is harmless because the factor is constant; for varying lengths it changes both the projection axes and the membership updates. So Algorithm 1 is not coordinate descent on the stated objective, and the convergence check has no formal basis. The variable-length simulations in Scenario 3 are run under this mismatched algorithm, so the paper's advertised ``varying lengths'' feature does not actually solve the problem it poses.\n\nCredit where it's due: the paper is clearly written, the simulation design is reasonable (VARMA and EEG mixtures), and the authors compare against VPCA and FCMD. The EEG drowsiness application is the best part: for subject 11, the CVI-optimal m=1.1 aligns with the highest RI, and the membership matrix reveals transitional states that hard clustering would miss. Code is on GitHub.\n\nSmaller soft spots: the simulation tables report means without standard deviations; FCPCA uses a single random initialization for the RI results even though Algorithm 1 recommends multiple starts; fuzzy-series detection is near zero at low dimensions (Table 3); there is an unresolved ``see Section??'' placeholder; and the CVI-based selection of m is validated on only one or two subjects. The abstract's claim that FCPCA ``outperforms several existing approaches'' is broader than the two fuzzy benchmarks support, and on some hard-clustering datasets VPCA beats it.\n\nThe algebraic gap is the main issue. It is fixable: either include the (T_i - l) weight in Eq. (8) and rerun the variable-length experiments, or change the objective to match what the algorithm actually optimizes. As written, the core varying-length claim does not hold. This paper deserves a serious referee, but I would send it back for major revision. For someone working on fuzzy time-series clustering, the fixed-length version is a useful read; the variable-length part should not be cited until the mismatch is resolved.","headline":"FCPCA is a sensible fuzzy extension of CPCA clustering, but the varying-length version does not minimize the stated objective; the paper needs a fix or a revised claim.","tokens_in":16824,"tokens_out":2929,"would_cite":false,"duration_ms":27875,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62M10","62H25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces FCPCA, a fuzzy clustering method for high-dimensional multivariate time series that assigns partial memberships from reconstruction error in cluster-specific common principal component subspaces, and reports that it…","keywords":["fuzzy clustering","multivariate time series","common principal component analysis","reconstruction error","EEG signals","fuzziness parameter","cluster validity index","high-dimensional time series"],"falsifier":"Simulate two clusters with identical lag-1 and lag-2 cross-covariance blocks but different lag-3 blocks, run FCPCA with $L=2$ as the paper does, and check whether the Rand index of the hard labels stays at chance; if it does, the reconstruction error is blind to the true cluster structure.","tokens_in":15766,"feed_emoji":"🧠","tokens_out":5546,"duration_ms":54461,"temperature":0.7,"pith_summary":"The paper proposes a fuzzy clustering method, FCPCA, for multivariate time series that can be high-dimensional and of unequal lengths. Instead of measuring distance to a centroid, FCPCA measures how well each series is reconstructed after projection onto a low-dimensional common subspace shared by each cluster. The authors argue this reconstruction-based objective handles the curse of dimensionality, that a data-driven criterion selects the fuzziness parameter and number of clusters objectively, and that partial memberships reveal transitional states that hard clustering misses. Simulations and real EEG data are used to support the claim that FCPCA outperforms several existing fuzzy and hard clustering approaches.","feed_headline":"Fuzzy clustering outmatches hard labels on high-dimensional time series","feed_subtitle":"FCPCA projects each cluster onto its own subspace and flags samples caught between drowsy and alert.","key_machinery":"The mechanism is the fuzzy reconstruction-error objective: minimize $\\sum_{i,s} u_{is}^m \\sum_l \\|\\hat{X}_i(l) - \\hat{X}_i(l) C_s(l) C_s(l)^\\top\\|^2$, where $u_{is}$ is the fuzzy membership of series $i$ in cluster $s$, and $C_s(l)$ is the projection subspace for cluster $s$ at lag $l$. Each subspace comes from the SVD of the membership-weighted block covariance matrix $\\tilde{\\Sigma}_s(l) = \\frac{\\sum_i u_{is}^m \\hat{\\Gamma}_i(l)}{\\sum_i u_{is}^m}$, where $\\hat{\\Gamma}_i(l)$ stacks the lag-0 and lag-$l$ cross-covariance estimates into a $2p \\times 2p$ block. Membership updates make $u_{is}$ inversely proportional to reconstruction error, and a modified Xie-Beni index selects the fuzziness parameter $m$ and number of clusters $S$. The common-principal-component idea is that all series in a cluster share a low-dimensional linear subspace of lagged covariance space, so each series is best represented by its own cluster's projection axes.","core_discovery":"The central claim is that fuzzy clustering of multivariate time series should be driven by reconstruction error in cluster-specific common subspaces rather than by distance to a representative. FCPCA builds, for each cluster and lag, a membership-weighted block cross-covariance matrix, extracts its leading principal components as the cluster's projection axes, and alternates between updating those axes and updating each series' membership degrees from the ratio of reconstruction errors. The paper reports that this yields superior clustering accuracy on VARMA-generated series with varying lengths and dimensions, that it successfully identifies genuinely mixed or transitional series, and that on EEG driver-drowsiness data it produces memberships revealing alert-to-drowsy transition states that hard clustering cannot express.","pith_inferences":["Editorial inference: If reconstruction error in a cluster-specific subspace is the right criterion, a natural extension is to handle missing or irregularly sampled observations by estimating the block covariance matrices from pairwise-complete lags; the paper does not treat missingness.","Editorial inference: The membership trajectory over time could serve as a continuous drowsiness score in driver monitoring, since partial memberships track the approach to a drowsy state rather than forcing a binary label; the paper mentions this idea but does not test it as a forecasting signal.","Editorial inference: Because FCPCA assumes each cluster is a single low-dimensional linear subspace, a cluster containing multiple distinct regimes (for example, alternating attentive and fatigued stretches within one subject) may need to be split into more clusters than the number of true states, making the validity index's choice of $S$ critical.","Editorial inference: The use of lagged block covariance matrices captures only second-order temporal structure, so clusters that differ only in higher-order moments or nonlinear dynamics would be invisible to FCPCA, and testing that boundary would clarify the method's scope."],"forward_implications":["FCPCA gives competitive fuzzy clustering accuracy on VARMA-generated multivariate time series across lengths 200–600 and dimensions 20–100, with mean Rand indices often at 0.91 or above for moderate fuzziness values, beating VPCA and fuzzy C-medoids.","The method can flag mixed or transitional series: VARMA(1,1) series that combine VAR(1) and VMA(1) behavior receive low membership in both pure clusters, so fuzzy memberships act as an explicit indicator of ambiguous observations.","On EEG driver-drowsiness data, FCPCA produces memberships that locate transitional alert-to-drowsy states within individual subjects, a capability hard clustering lacks; for subject 11, the automatically selected fuzziness parameter also yields the highest Rand index among the values tested.","Even a low fuzziness setting, $m=1.1$, improves the Rand index over the hard CPCA baselines on several real multivariate time series datasets, suggesting that mild fuzziness absorbs labeling uncertainty without sacrificing crisp interpretability.","The iterative scheme only requires SVD of $S$ membership-weighted $2p \\times 2p$ matrices per iteration, which keeps the method computationally feasible for moderately high-dimensional data; the paper reports consistent performance up to 128 EEG channels.","A data-driven modified Xie-Beni index selects both the fuzziness parameter and the number of clusters, removing two user choices that are usually made subjectively."],"supporting_citations":[{"why":"Supplies the original CPCA-based hard clustering method Mc2PCA, which FCPCA extends to a fuzzy setting.","marker":"Li (2019)"},{"why":"Provides the ROBCPCA robust CPCA variant whose temporal-dependence modifications the hard-clustering baseline in this paper adopts.","marker":"Ma et al. (2024)"},{"why":"Defines the classical fuzzy C-means framework whose fuzziness parameter and membership constraints FCPCA borrows.","marker":"Bezdek et al. (1984)"},{"why":"Source of the Xie-Beni cluster validity index that the paper adapts for data-driven selection of $m$ and $S$.","marker":"Xie and Beni (1991)"},{"why":"Describes the VPCA fuzzy clustering baseline that FCPCA is compared against in the simulation study.","marker":"He and Tan (2018)"},{"why":"Defines the fuzzy C-medoids clustering model used as the FCMD baseline in the numerical comparisons.","marker":"Coppi et al. (2006)"},{"why":"Introduces the Rand index used to evaluate the agreement between the crisp version of FCPCA's labels and the true partitions.","marker":"Rand (1971)"},{"why":"Supplies the AR(2) mixture model used to simulate EEG-like multivariate time series in Section 4.2.","marker":"Ombao and Pinto (2024)"},{"why":"Provides the EEG driver-drowsiness dataset analyzed in Section 5 and its experimental protocol.","marker":"Cui et al. (2022)"}],"fun_headline_variants":["Fuzzy clustering beats crisp clustering on high-dim time series","FCPCA: fuzzy clustering with cluster-specific common subspaces","Shared PCA axes improve fuzzy clustering of multivariate series","Fuzzy memberships uncover drowsy-alert transitions in driving EEG","New fuzzy method scales to high-dim time series of varying lengths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes each cluster of multivariate time series is second-order stationary and that its structure lives in a low-dimensional linear subspace of lagged cross-covariance blocks; if clusters are non-stationary or not linearly separable in that covariance space, the reconstruction errors that define memberships will not separate them.","fun_headline_variants_meta":{"raw":{"variants":["Fuzzy clustering beats crisp clustering on high-dim time series","FCPCA: fuzzy clustering with cluster-specific common subspaces","Shared PCA axes improve fuzzy clustering of multivariate series","Fuzzy memberships uncover drowsy-alert transitions in driving EEG","New fuzzy method scales to high-dim time series of varying lengths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000798,"raw_usage":{"total_tokens":3484,"prompt_tokens":891,"completion_tokens":2593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2510}},"tokens_in":507,"tokens_out":2593,"duration_ms":18299,"temperature":1.0,"reasoning_tokens":2510,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:20:58.554113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate two clusters with identical lag-1 and lag-2 cross-covariance blocks but different lag-3 blocks, run FCPCA with $L=2$ as the paper does, and check whether the Rand index of the hard labels stays at chance; if it does, the reconstruction error is blind to the true cluster structure.","supporting_citations":[{"cited_title":"C., Ehrlich, R., and Full, W","cited_arxiv_id":null,"evidence_quote":"Defines the classical fuzzy C-means framework whose fuzziness parameter and membership constraints FCPCA borrows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the Xie-Beni cluster validity index that the paper adapts for data-driven selection of $m$ and $S$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the fuzzy C-medoids clustering model used as the FCMD baseline in the numerical comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Rand index used to evaluate the agreement between the crisp version of FCPCA's labels and the true partitions."},{"cited_title":"and Pinto, M","cited_arxiv_id":null,"evidence_quote":"Supplies the AR(2) mixture model used to simulate EEG-like multivariate time series in Section 4.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EEG driver-drowsiness dataset analyzed in Section 5 and its experimental protocol."}],"review_version":1}