acceptodds
Under review as a conference paper at ICLR 2027

Multi-View Time-Series Clustering with Temporal Structure and Tensor Regularization

Abstract

We propose a tensor-based method for multi-view time-series clustering that jointly models temporal information and structure shared across views. Within each view, we combine normalized dissimilarities derived from trajectory states and first-order differences to construct a clustering objective incorporating pairwise relations and triplet geometry. View weights are learned adaptively to balance the view-specific fitting losses. Across views, we learn a common clustering embedding through symmetrized products of subspace projectors and impose tensor nuclear norm regularization on orthogonally aligned view-specific embeddings. Orthogonal alignment accommodates differences in embedding bases while preserving the subspace of each view. The final cluster assignments are obtained from the common embedding. To solve the resulting nonconvex, nonsmooth problem, we develop a two-level algorithm combining proximal majorization–minimization with an inexact augmented Lagrangian method. Under the stated conditions, we prove that the splitting constraint residual converges to zero and every accumulation point of the outer primal iterates is a first-order stationary point of the original problem. Comparative experiments on multiple datasets demonstrate the method’s effectiveness in clustering multi-view time series. Parameter sensitivity analyses and ablation studies further assess the effects of key parameters and individual components on clustering performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Multi-view clustering, Time-series clustering, Tensor regularization, Subspace consensus, Higher-order relations

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zvJmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DTMVC for multi-view time series clustering. The method combines normalized state distances with first-order difference distances to construct fixed pairwise relationships and a sparse ternary geometric loss, thereby adaptively learning orthogonal embeddings for each view. The cross-view component combines projected product consensus with tensor norm regularization of orthogonally aligned embeddings, and finally runs k-means on the normalized H-row common embedding. The authors propose a two-layer proximal MM–iALM algorithm and prove, under specified conditions, that the splitting residuals tend to zero and that the clusters of the original variables exhibit first-order fixed-point properties. Comparisons across six datasets, parameter analysis, and ablation studies yield positive results; however, the experimental report and implementation details still require clarification.

Critical strengths

1.The model addresses coordinate discrepancies between temporal geometry and embedding representations. States and first-order differences are combined to construct pairwise and triplet sample relationships; projection–product consensus is combined with tensor norm regularization of orthogonally aligned embeddings to reduce the impact of coordinate selection on tensor constraints. 2.The appendix provides a detailed explanation of the update mechanism and convergence proofs. Table 2 reports the best values for all metrics across six datasets using the same set of parameters; these results support the method’s potential practical value.

Critical weaknesses

1.The lack of consistency in the experiments and the model selection protocol make it difficult to fully assess the performance evidence. DSADS achieves an ACC of 70.37 ± 4.47 in Table 2, while the Full DTMVC model—claimed to be the optimal result from the grid search in Table 3—achieves 68.38 ± 4.99; the parameter combinations in Table 2 also fall within that grid. Differences may be attributed to variations in random runs or settings, but this has not been explained. It remains to be clarified whether the parameter selection for the grid search was based on labeled data, and the method for selecting parameters during unlabeled deployment also requires further explanation. 2.The dataset sizes in Table 1 range from 312 to 9,120 samples; the performance of the method on larger datasets has not yet been demonstrated. The time complexity analysis in the main text primarily covers pairwise matrix multiplication and sparse ternary terms; the appendix also does not provide additional details on the computational costs of steps such as orthogonal updates and tensor FFT/SVD, nor does it address the impact of the number of inner and outer iterations on the total runtime. Furthermore, the paper lacks a comparison of runtime with baseline methods. We recommend supplementing the paper with a comprehensive time complexity analysis and reporting runtime results across a range of sample sizes to validate the method’s computational efficiency and scalability.

Questions for authors
  1. Please explain the differences in DSADS results between Table 2 and Table 3. Do both tables use the same data processing, optimization initialization, and k-means random seeds? Why are the optimal results from grid search lower than those from fixed parameter configurations within the grid?

  2. How were the uniform parameters in Table 2 determined? Were the true labels from the evaluation dataset used? For practical unlabeled scenarios, how should the three hyperparameters be selected?

  3. Did the comparison methods use the parameters recommended in the original paper, default parameters, or were they fine-tuned for each dataset? If fine-tuning was performed, please specify the parameter search range to ensure a fair comparison.

  4. Could you provide a complete time complexity analysis covering the main update steps and the number of inner and outer iterations? Additionally, please report the comparison of runtime and clustering performance with baseline methods as the sample size gradually increases to verify the method’s scalability?

1 comment
AuthorsOpenReview ↗

Response to Reviewer zvJm

Thank you for your careful review. We appreciate your pointing out the discrepancy between the two DSADS results and asking us to clarify the parameter-selection protocol and computational cost. We address each point below.

1. DSADS results in Table 2 and Figure 2. Thank you for pointing out this discrepancy. The DSADS run reported in the original Table 2 was continued until convergence. The parameter sensitivity analysis in Figure 2 used a fixed budget of 50 iterations. The highest ACC in Figure 2 was therefore specific to that budget. Even at a parameter setting included in the grid, continuing the optimization beyond 50 iterations could yield a different embedding and a higher ACC. We should have used the same stopping rule in both experiments and stated it explicitly.

2. Parameter selection and the role of labels. The original manuscript did not document the rationale for the fixed settings clearly enough. In particular, the archived DSADS fit used γ=1\gamma=1, which differs from the uniform setting described in the original text. Section 5.1 and Section 6 of the supplement now specify the revised protocol. We calculate a dataset-specific reference β\beta from pairwise and triadic reference losses using only the input geometry. With β\beta fixed, we evaluate all 64 (γ,λ)(\gamma,\lambda) combinations and pool the results from 50 kk-means starts per combination. Ground-truth labels are used to score the clustering outcomes, and every combination contributes to the reported statistics.

Table 3 serves a different purpose. Its Full DTMVC entry uses the highest mean ACC in the sensitivity grid, so that selection uses the benchmark labels. For unlabeled deployment, the reference β\beta remains applicable. Selecting a single (γ,λ)(\gamma,\lambda) pair would require a separate label-free criterion, such as clustering stability across resamples. We have not evaluated such a selection rule here.

3. Baseline parameter settings. We reviewed the parameter sources and input adaptations for all baselines and expanded the comparison to include MUSLA and CSMVC. We prioritize settings recommended in the original papers or released implementations. Where a sensitivity study supplies candidate ranges, we use those ranges to inform the grid and include every evaluated combination in the reported statistics. Section 5 of the supplement documents the settings for each method. For example, the S²MVTC, DSTL, and LMTC grids draw on the respective papers' sensitivity studies, and EMTC uses a configuration from its released training script. All methods receive the benchmark cluster count.

4. Computational cost and sample-size scaling. Section 4 of the supplement now gives costs for the main updates, including the orthogonal polar updates, the reduced eigensystem for the common embedding, and the FFT and slice SVDs in the tensor update. If tst_s is the number of inner sweeps in outer iteration ss, the sweep-dependent cost scales with I=∑stsI=\sum_s t_s, with I≤SKinI\leq S K_{\mathrm{in}} under a cap of KinK_{\mathrm{in}} inner sweeps over SS outer iterations. The analysis also accounts for one-time geometry construction and the final kk-means readout.

Section 8 adds an eight-size experiment with three random repeats per size on MHEALTH and PASTIS-R. Mean runtime increased from 19.36 to 120.19 seconds on MHEALTH and from 5.70 to 48.60 seconds on PASTIS-R. Mean ACC stayed between 54.52% and 56.49% on MHEALTH. On PASTIS-R, it rose from 90.98% at 250 samples to 95.45% at 750 samples and then remained close to that level. Thus, across the tested sizes, the increase in runtime was accompanied by stable or improved mean ACC overall. These measurements characterize DTMVC on the two datasets. A head-to-head comparison with baselines at the same sample sizes remains an important follow-up.

Reviewer_egpfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DTMVC for multi-view time series clustering. In each view, it builds dissimilarities from trajectory states and first-order differences, and uses them in a pairwise loss and a triadic loss with adaptive view weights. Across views, it learns a common embedding through projection products and adds a tensor nuclear norm on orthogonally aligned view embeddings. A two-level MM–iALM algorithm is given with a convergence proof. Experiments on six datasets show better results than the compared baselines.

Critical strengths
  1. The tensor nuclear norm on aligned embeddings is a simple and reasonable idea. It makes the regularizer independent of the basis of each view.
  2. The paper is clearly and the convergence analysis is complete.
Critical weaknesses
  1. The temporal modeling contribution is thin. Everything related to time is contained in Eq. (1): a per-feature RMS distance over time-aligned points plus a first-difference term. All trajectories must share a temporal grid within each view, and ∆_v and W_v are computed once before optimization and then fixed (end of Sec. 3.1, App. A.4). After this preprocessing step, the model is identical to a static multi-view spectral clustering method. The current mean-pooling ablation is a weak control. A more informative one would construct ∆_v from distances on flattened trajectories and run the same model, which would directly answer whether the first-difference term adds anything beyond simple flattening.
  2. Two highly relevant works on multi-view multivariate time-series clustering are neither cited nor compared: Zhang et al., "Multiview Unsupervised Shapelet Learning for Multivariate Time Series Clustering" (TPAMI 2023), and He et al., "CSMVC: A Multiview Method for Multivariate Time-Series Clustering" (TCYB 2022). Both address the same setting and should be discussed and included as baselines; currently MVCIMTS is treated as the only method of this kind.
Questions for authors
  1. The novelty narrative rests on "temporal + multi-view", but the ablation does not consistently support the temporal part. In Table 3, w/o Differences exceeds the full model on Tactile12 in NMI and ARI, the losses on DSADS and HPC-ODA fall within the reported standard deviations, and on NinaPro the Mean-pooling variant, which discards temporal information entirely, is nearly identical to the full model. The only large gap appears on PASTIS-R, which happens to have the shortest sequences (T_v = 12). Could the authors clarify under what conditions the temporal terms actually matter?
  2. In Table 2, all 24 entries of the MVCIMTS+K-means column have means that are multiples of 0.5 and standard deviations that are square roots of integers (e.g., 3.16 = √10, 7.07 = √50, 7.75 = √60). For datasets with thousands of samples and continuous metrics such as NMI and ARI, it is very unlikely that the mean and SD of ten runs would all fall on such a grid. Could the authors explain how these numbers were obtained and provide the raw per-run results?
2 comments
AuthorsOpenReview ↗

Response to Reviewer egpf

Thank you for carefully reviewing our modeling assumptions, experimental controls, and reported results. We appreciate you drawing our attention to the irregular values in the original comparison. We should have caught this issue when compiling the results. Your suggestion to use flattened trajectories as a control also helped us clarify when temporal changes contribute to clustering.

We reviewed the parameter settings, input adaptations, and evaluation procedures for all baselines and updated the main comparison. Where the original papers or released implementations recommend a configuration, we prioritize that setting. Where a paper provides a sensitivity study without recommending a configuration for our setting, we use its reported ranges to define a candidate grid, evaluate every combination in that grid, and weight the combinations equally in the reported statistics. Ground-truth labels are reserved for evaluation. Several entries in Table 2 have changed as a result. Section 5 of the supplement documents the settings for each method.

W1. How temporal information enters the model and what the flattened-trajectory control tests. The idea behind DTMVC is to represent temporal variation through fixed relationships between samples. We first construct these relationships from trajectory levels and changes between consecutive observations, then use them for multi-view clustering. Within each view, trajectories share a time grid. The state distance compares observation levels at corresponding times, and the first-difference distance compares signed changes over corresponding intervals. For example, two trajectories can have different level offsets and still rise and fall in similar ways at the same times. First differences capture this similarity in their changes.

The pairwise matrix records dissimilarities between samples. The third-order tensor represents the shapes of local triples through the angles of their comparison triangles. We also considered four-sample relationships during the design stage, such as tetrahedral geometry in an appropriate Euclidean setting. We chose sparse triples because triangle angles provide an interpretable description of local shape and keep the computational cost manageable. Temporal information thus enters through the construction of the pairwise and triadic relationships. The optimizer uses both sets of relationships to learn a common clustering embedding.

We agree that the flattened-trajectory control is more direct for testing the contribution of first differences. Mean pooling retains the temporal average of each feature. The flattened control retains the observations at each time point. In Section 7 of the supplement, ϑ=0\vartheta=0 corresponds to this control under our distance convention. For each feature, we compute the Euclidean distance between the aligned trajectories, scaled by 1/Tv1/\sqrt{T_v}. We sum these distances across features and normalize by the mean off-diagonal distance within the view. We then rebuild the pairwise and triadic relationships and refit the same model with the remaining settings fixed. This comparison isolates the effect of adding first differences to the featurewise state-distance construction.

W2. Related work and additional baselines. Thank you for pointing us to MUSLA and CSMVC. The revised manuscript cites and discusses both methods and includes them in Table 2. MUSLA constructs views from shapelets of different lengths, and CSMVC models shared and view-specific structure across its constructed views. DTMVC starts from the sensor or feature groups supplied with the data and builds pairwise and triadic temporal relationships within each view. Section 5 of the supplement documents the input organization, parameter sources, and clustering procedures for both added baselines.

AuthorsOpenReview ↗

Response to Reviewer egpf

Q1. When first differences improve clustering. As you observed, removing first differences improved some metrics on Tactile12 in the original Table 3. Some of the differences on DSADS and HPC-ODA were also small. We therefore evaluate the first-difference weight separately. Section 5.5 fixes ϑ=0.5\vartheta=0.5 for the remaining ablations, and Section 7 of the supplement compares seven weights, including the state-only and first-difference-only endpoints. This supplementary analysis focuses on ACC.

Five of the six datasets attain their highest mean ACC at a mixture of the two distances on the tested grid. On PASTIS-R, the equal mixture reaches 95.80%, compared with 90.43% for state distances alone and 93.68% for first differences alone. Tactile12 and HPC-ODA attain their highest mean ACC at ϑ=0.3\vartheta=0.3. On DSADS, the means at ϑ=0\vartheta=0 and 0.10.1 differ by only 0.19 percentage points. On NinaPro, state distances alone give the highest mean ACC, and ACC decreases as the first-difference weight increases. These results support a dataset-dependent role for first differences.

A larger first-difference weight can improve separation when samples from the same class show similar changes despite shifts in their observation levels, and samples from different classes change in different ways. A smaller weight preserves more of the signal when classes are distinguished mainly by their levels at corresponding times. PASTIS-R has 12 observations per view, which provide 11 intervals over which to compare changes. Its results illustrate that useful temporal information can be present even in a short sequence. The comparison with mean pooling captures the broader effect of discarding temporal structure. The weight sweep directly tests the contribution of first differences.

Section 8 of the supplement also adds a sample-size study on MHEALTH and PASTIS-R. We evaluate eight sample sizes with three random repeats per size. Runtime increases with sample size on both datasets. Mean ACC remains broadly stable on MHEALTH. On PASTIS-R, it improves at smaller sample sizes and then levels off.

Q2. Irregular MVCIMTS+K-means statistics and run-level records. Thank you for pointing out that several means were concentrated at multiples of 0.5 and that the standard deviations followed a regular pattern. Our review confirmed an error in the reporting of the affected entries. For HPC-ODA, we recovered the original run-level scores and predicted labels and reproduced the reported means and sample standard deviations. For the other affected entries, we could not recover run-level records that reproduce the published summaries. The specific cause of the numerical pattern therefore remains unresolved. We have withdrawn the unverifiable earlier values in the revision.

We also clarify the evaluation protocol for the MVCIMTS adaptation. We treat each multichannel sensor group as a view, concatenate the normalized view embeddings, and apply kk-means to obtain the final assignments. Section 5 of the supplement describes the parameter grid, clustering starts, and aggregation procedure. For verification, the recoverable original run records and the new run records used in the revision should be identified separately. Each verifiable mean and sample standard deviation can then be checked directly against the corresponding run-level scores.

Reviewer_38ozOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DTMVC, a dynamic multi-view clustering method for sample-matched multi-view trajectories. The method constructs pairwise and sparse triadic clustering losses from normalized state and first-difference dissimilarities, couples view subspaces through a projection-product consensus term, and regularizes orthogonally aligned embeddings with a tensor nuclear norm. A two-level proximal MM-iALM algorithm is developed, and a convergence result is established showing that splitting residuals vanish and every primal accumulation point satisfies split KKT conditions and corresponds to a first-order stationary point of the original problem. Experiments on six real-world datasets report consistent improvements over six baselines.

Critical strengths
  1. The formulation is carefully motivated. The distinction between discrepancies in observation levels and those in adjacent changes is a sensible modeling choice for trajectory data, and the triadic loss based on comparison-triangle angles provides a shape-aware, scale-invariant geometric signal that is not captured by pairwise distances alone. The use of sparse triplet lists rather than dense third-order tensors is a practical and well-justified implementation decision.

  2. The empirical evaluation is broad, covering six datasets from different domains, five metrics, parameter sensitivity over a 512-configuration grid, and a seven-variant ablation that isolates each component (differences, triadic loss, projection product, tensor term, alignment, single-sweep, mean-pooling).

Critical weaknesses
  1. The convergence guarantee, while carefully stated, rests on assumptions that are somewhat stronger than the algorithm as implemented. Theorem 1 requires that "all prescribed MM and proximal block updates are computed exactly," yet Algorithm 1 caps the inner loop at 50 iterations and uses tolerance-based acceptance. The paper should clarify whether the guarantee applies to the practical algorithm or only to an idealized variant, and whether the finite-inner-termination lemma remains meaningful when the cap binds.

  2. The experimental comparison is somewhat narrow relative to the framing. The paper positions itself against both static multi-view methods and time-series representation methods, but the baselines are adapted in nontrivial ways (flattening sequences, concatenating channels, encoding unequal-length views separately). Since these adaptations may disadvantage the baselines, a discussion of adaptation fairness, ideally with a baseline run under its native setting on at least one dataset, would be helpful.

Questions for authors
  1. In Eq. (4), the triadic loss uses L_{3,v}(F_v) = sqrt(c) * ||W_v||F - Phi{3,v}(F_v), and the paper notes that both constants are retained because they affect the optimization of alpha. Could the authors clarify how the shifts b_{2,v} and sqrt(c) * ||W_v||_F interact with the weight update in Eq. (12)? In particular, when a view's shifted energy is near zero, the zero-energy rule assigns uniform weight over the zero set; does this rule remain stable if the shifts are large relative to the data-dependent terms?

  2. The projection-product consensus M_x averages over all view pairs with equal weight, while adaptive weights alpha_v apply only to the fitting energies. What is the rationale for not letting the consensus term inherit the adaptive weights? An ablation or discussion of whether weighted consensus would improve robustness to low-quality views would be valuable.

  3. The paper fixes beta = 1, gamma = 10, and lambda = 0.01 across all six datasets in Table 2, while the sensitivity study shows that preferred beta varies considerably (e.g., HPC-ODA varies by 19.00 percentage points). How were these fixed values selected, on a held-out split, or by best average performance? Clarifying this would help assess whether the reported gains reflect a fair protocol.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 38oz

Thank you for your careful reading and constructive feedback. Your comments on the scope of the convergence guarantee and the fairness of the baseline comparisons helped us identify where the original manuscript needed clarification.

The original evaluation used different parameter-selection conventions across baselines. In the revision, we prioritize settings explicitly recommended in the original papers or the authors' public implementations. When a paper reports a sensitivity study and gives no recommended setting for our benchmark, we use the reported ranges to define a candidate grid, evaluate every combination in that grid, and give each configuration equal weight in the reported statistics. We reserve ground-truth labels for evaluation and report means and sample standard deviations across the evaluated configurations and random starts. Section 5 of the supplement documents the parameter sources, input adaptations, and aggregation rules for each baseline.

We also added Section 5.3 to compare three inner-solver strategies. We address your specific comments below.

W1. Scope of Theorem 1 and the 50-sweep cap. Revised Algorithm 1 makes the inner acceptance condition explicit, Sst≤εs\mathcal S_s^t\leq\varepsilon_s. At each outer iteration, we hold the multiplier and penalty parameter fixed while updating the blocks in sequence. We update the multiplier after accepting the inner point. The exactness assumption in Theorem 1 concerns each prescribed MM or proximal block update. The inner subproblem as a whole is solved to the prescribed stationarity tolerance.

The finite-termination lemma in Section 2 of the supplement guarantees that, under the theorem's assumptions, continued inner updates reach the positive tolerance for the current outer iteration after finitely many sweeps. The required number of sweeps can vary across datasets and outer iterations and can exceed 50. If the residual satisfies the tolerance at sweep 50, the inner point meets the acceptance condition. If the residual remains above tolerance when the cap is reached, accepting that point is a finite-budget choice outside the scope of the theorem's acceptance condition. Section 5.1 now states the budget used for the main comparison, and Algorithm 1 specifies the residual-based rule analyzed in Theorem 1.

Section 5.3 compares No-inner, Fixed-50, and residual-based stopping. Figure 2 uses illustrative trajectories to show the objective and ACC over outer iterations, with three-dimensional insets showing how the inner objective evolves across outer iterations. The cumulative number of inner sweeps, I=∑stsI=\sum_s t_s, accounts for the additional inner work in the complexity expression in the main text and its derivation in Section 4 of the supplement.

W2. Baseline adaptations and fairness. We expanded the comparison from six to eight baselines by adding MUSLA and CSMVC. All methods receive the benchmark cluster count. For static multi-view methods, we vectorize trajectories within each view and preserve sample correspondence across views. For methods that take a single multivariate series, we specify how the channels are aligned and concatenated. Our MVCIMTS adaptation treats each sensor group as one view and applies kk-means to the concatenated, normalized view embeddings for the final assignments.

Section 5 of the supplement also documents the sources of the parameter settings. For example, the S²MVTC, DSTL, and LMTC grids draw on the respective papers' parameter studies. EMTC uses a configuration from the released training script, and MUSLA uses the initialization from the authors' example. For CSMVC, we give the grid used in our evaluation and include all 64 combinations in the reported statistics. The supplement identifies the input and implementation choices we made for this benchmark.

We appreciate your suggestion to include a native-setting control. A matched comparison between native and adapted settings would help quantify the effect of the adaptations. The present revision expands the baseline set and makes the adaptations and their scope explicit.

AuthorsOpenReview ↗

Response to Reviewer 38oz

Q1. Energy shifts and the stability of the view weights. Let Ev=L2,v+βL3,vE_v=\mathcal L_{2,v}+\beta\mathcal L_{3,v} denote the fitting energy of view vv. The pairwise loss subtracts b2,vb_{2,v}, the minimum of the relaxed pairwise criterion under the orthogonality constraint. The triadic loss uses the upper bound c∥Wv∥F\sqrt c\lVert\mathcal W_v\rVert_F minus the actual triadic score. Both constructions give nonnegative losses. The shifts are constant with respect to the embedding during its update. In the weight update, they enter through the energies EvE_v and affect the ratios across views. When all energies are positive, the weight update gives

=\left(\frac{E_u}{E_v}\right)^{1/(r-1)}.$$ When some energies are exactly zero, any simplex weight vector supported on the zero-energy views minimizes the weight subproblem. We choose uniform weights on this set as a deterministic convention. The ratio formula still applies to small positive energies. If one energy approaches zero while the others stay bounded away from zero, its weight approaches one. If several energies approach zero together, their limiting weights depend on their relative rates of decay. The sensitivity therefore depends on relative errors in the shifted energies. The update can remain stable with large shifts if the resulting energies stay well away from zero. If two large quantities nearly cancel, even small absolute errors in their difference can lead to sizable changes in the weights. The zero-energy rule specifies a minimizer at exact zeros, and the ratio formula describes the behavior near zero. **Q2. Equal weighting in projection-product consensus.** The adaptive weight $\alpha_v$ controls the contribution of view $v$ to the pairwise and triadic fitting terms. The projection-product consensus term measures agreement between pairs of view subspaces through the common embedding. Equal weighting lets every pair contribute, including pairs that involve a view with higher individual fitting energy that may still provide complementary information. Using pair weights proportional to $\alpha_u\alpha_v$ would reduce both the fitting contribution of a low-weight view and the contribution of all pairs involving that view. This could reduce the influence of a noisy view, and it could also suppress complementary information. In the limiting case where all weight falls on one view, every cross-view product $\alpha_u\alpha_v$ vanishes. This change would also introduce $\boldsymbol\alpha$ into the consensus term and require a new weight update. The projection-product ablation in Table 3 assesses the effect of the consensus structure itself. We appreciate your suggestion to examine weighted consensus. A direct robustness test would compare the two weighting schemes while progressively degrading one view and holding the remaining settings fixed. **Q3. Parameter selection for Table 2.** The original Table 2 used $(\beta,\gamma,\lambda)=(1,10,0.01)$ on all six datasets. The rationale for this shared setting was left implicit in the original manuscript. Your observation about the sensitivity results highlighted why the choice deserves a clearer protocol. In the revision, we determine a dataset-specific reference $\beta$ from the input and evaluate the other two parameters over a grid. As described in Section 6 of the supplement, we construct a reference embedding from the pairwise distances and triadic tensors and evaluate the pairwise and triadic reference losses, $s_2$ and $s_3$, at that embedding. Dividing each loss by its reference value gives a relative coefficient of $s_2/s_3$ for the triadic term when the pairwise coefficient is set to one. We choose the candidate $\beta$ closest to this ratio on the logarithmic scale. This procedure uses only the input geometry. With the reference $\beta$ fixed, we evaluate all $8\times8=64$ combinations of $\gamma$ and $\lambda$. For each fitted embedding, we run 50 independently initialized $k$-means trials. Table 2 reports the mean and sample standard deviation of each metric, giving equal weight to all configurations and starts. Section 5.4 and Figure 3 retain the full three-parameter sensitivity study to show how performance varies across datasets and settings.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.