acceptodds
Under review as a conference paper at ICLR 2027

TIGER: Target-Aware Interventional Graph-Structure Decoupling and Retrieval for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to learn effective clustering representations from multi-view data with partially missing views. Although considerable progress has been made, existing methods still face several challenges: 1) Many approaches recover missing views through direct feature reconstruction or global information sharing, which may ignore the distribution characteristics of the target view and lead to inaccurate representation generation. 2) The structural relationships among samples and the complementary information across views are insufficiently explored during missing-view recovery, making the learned representations sensitive to incomplete and noisy observations. 3) Existing methods usually lack an explicit mechanism to evaluate the reliability of recovered information, causing uncertain completion results to negatively affect representation fusion and clustering optimization. To address these issues, we propose TIGER: Target-aware Interventional Graph-structure Decoupling and Retrieval for Incomplete Multi-view Clustering. Specifically, TIGER first learns view-specific latent representations and separates transferable and view-dependent information to construct more reliable cross-view relationships. Then, a target-aware retrieval strategy is developed to recover missing views by selecting informative anchors from observed target-view samples according to cross-view structural evidence. Furthermore, an intervention-based completion mechanism is introduced to identify and reduce unreliable local correlations, enabling robust representation recovery under severe missing conditions. Finally, the recovered representations are integrated with clustering learning through a unified optimization framework. Extensive experiments on seven benchmark datasets under various missing rates demonstrate that TIGER consistently outperforms state-of-the-art IMVC methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Target-Aware Retrieval, Graph-Structure Decoupling, Graph Intervention

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_7hhdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a Target-aware Interventional Graph-structure Decoupling and Retrieval (TIGER) for Incomplete Multi-view Clustering. It first learns view-specific latent representations with the goal to separate transferable and view-dependent information. Then, a retrieval strategy is developed to recover missing views by selecting informative anchors from observed target-view samples according to cross-view structural evidence. Then, an intervention-based completion mechanism is introduced to identify and reduce unreliable local correlations. Experiments on 7 benchmark datasets under 4 missing rates demonstrate the effectiveness of the proposed method.

Critical strengths

The proposed method is compared with the most recent baselines and the results verifies the effectiveness.

Critical weaknesses
  1. The method seems quite complicated. The reasons why each formula achieves the desired effect are not explained clearly, or the motivation behand each formula is not well illustrated.

  2. The symbol usage is not standardized. Some vectors are represented in bold (π\boldsymbol{\pi}), while others are not (sivs_i^v,pivp_i^v), which increases the difficulty of understanding.

Questions for authors
  1. For each missing proportion, how many independent repeated experiments were conducted?
  2. From Figure 7 of the supplementary materials, it seems that the clustering performance remains almost the same under different parameter settings. What could be the reason for this? Is it reasonable?
  3. How to understand targeted guided retrieval component? Is it necessary for the target view to be complete?
7 comments
AuthorsOpenReview ↗

W1. Explanation of the Methodology and Design Motivation

W1: We thank the reviewer for the careful evaluation and the opportunity to clarify the method's motivation. TIGER sequentially combines representation learning, graph-structure decoupling, target-aware retrieval, confidence estimation, and intervention-based completion.

Equations (1) and (2) establish view-specific representation learning. The encoder maps an observed feature to ziv=Ev(xiv)z_i^v=E^v(x_i^v), and the decoder reconstructs it as x^iv=Dv(ziv)\hat{x}_i^v=D^v(z_i^v). This preserves observed information for structural learning. Reconstruction uses observed entries only, and Smooth L1 limits the influence of unusually large residuals.

Equation (3) decomposes each latent representation into stable and private components, siv=Fsv(ziv)s_i^v=F_s^v(z_i^v) and piv=Fpv(ziv)p_i^v=F_p^v(z_i^v). The stable branch captures transferable relational information, whereas the private branch preserves view-dependent characteristics. It supplies complementary representations without forcing different views to share identical coordinates.

Equations (4) and (5) introduce graph-structure decoupling. Equation (4) constructs neighborhood distributions from cosine similarities between stable representations over jointly observed samples, with Ivu=Iv∩IuI_{vu}=I_v\cap I_u. The probability πijvu\pi_{ij}^{vu} describes the relative neighborhood importance of sample jj for sample ii in view vv. View uu uses the same jointly observed indices, enabling comparison across heterogeneous feature spaces.

Equation (5) combines Jensen-Shannon divergence with stable-private decorrelation. The term JSD⁡(πivu,πiuv)\operatorname{JSD}(\pi_i^{vu},\pi_i^{uv}) aligns corresponding neighborhood distributions, while ∣cos⁡(siv,piv)∣|\operatorname{cos}(s_i^v,p_i^v)| discourages redundant information between the two branches. The mixture is mivu=(πivu+πiuv)/2m_i^{vu}=(\pi_i^{vu}+\pi_i^{uv})/2. The divergence is nonnegative and vanishes when the distributions coincide. TIGER therefore aligns neighborhoods, not point-wise features, while retaining view-specific information.

Equations (6) and (7) use stable structure for target-aware retrieval. When view vv is missing for sample ii, Eq. (6) aggregates stable representations from available source views and normalizes them to form query qivq_i^v. The query summarizes cross-view evidence; it is a retrieval descriptor, not an imputed feature. Equation (7) searches the observed target bank IvI_v and selects the top-kk candidates by cosine similarity. Source views determine the query, whereas candidates and their latent codes come from the target view. This keeps retrieval evidence target-specific.

Equations (8) and (9) construct two target anchors. Stable weights αijv\alpha_{ij}^v aggregate candidate target representations into rivr_i^v, estimating the missing sample's principal location in the target space. Private weights βijv\beta_{ij}^v aggregate candidate private representations into uivu_i^v, preserving view-specific variations that stable aggregation may smooth. The anchors provide complementary evidence.

Equation (10) estimates retrieval confidence through normalized entropy, civ=1+∑j∈Aivαijvlog⁡(αijv)/log⁡(k)c_i^v=1+\sum_{j\in A_i^v}\alpha_{ij}^v\log(\alpha_{ij}^v)/\log(k). Concentrated weights produce higher confidence, while diffuse weights indicate ambiguous retrieval evidence. It measures retrieval concentration, not calibrated correctness, and regulates private information.

Equation (11) applies two independent stochastic interventions to private candidate edges while keeping the stable anchor and stable weights fixed. Remaining private weights are renormalized, with at least one candidate retained in each pass. This tests dependence on fragile private relations. Equation (12) predicts the target representation as yiv,t=riv+civGv([riv∥uiv,t])y_i^{v,t}=r_i^v+c_i^vG^v([r_i^v\Vert u_i^{v,t}]). The stable anchor supplies the target reference, confidence controls the residual, and averaging gives the completed representation.

Equation (13) trains completion with pseudo-missing observations. An originally observed target feature is temporarily hidden, and its original latent representation serves as the self-supervised target. Reconstruction encourages recovery, while consistency constrains predictions across interventions. No cluster labels are required. Finally, Eq. (14) jointly optimizes reconstruction, graph-structure decoupling, and graph-intervention objectives. The compact objective is L=Lrec+λsLs+λgLgL=L_{rec}+\lambda_sL_s+\lambda_gL_g, where the two weights balance structural alignment and intervention-based completion within the unified framework.

At inference, intervention predictions are averaged, observed views retain their representations, and completed views receive confidence-dependent weights before normalized fusion and K-means. Thus, reconstruction preserves observations, graph decoupling transfers relations, retrieval identifies target evidence, entropy measures confidence, and interventions improve robustness. The revised method description and Appendix B.5 detail the sequence and implementation.

AuthorsOpenReview ↗

Q1. Number of Independent Repeated Runs

Q1: We sincerely appreciate the reviewer’s attention to the experimental reproducibility protocol. Every dataset and every missing-rate setting is evaluated through five independent runs using five different random seeds. These are separate training and evaluation processes, rather than repeated evaluations of a single trained model. The reported results are the arithmetic mean and standard deviation calculated over the five independent runs for each method, dataset, missing rate, and evaluation metric.

For a prescribed missing rate rr, the observation indicator is generated according to Miv∼Bernoulli⁡(1−r)M_i^v\sim\operatorname{Bernoulli}(1-r). Each sample-view entry is independently retained with probability 1−r1-r and removed with probability rr. If a sample has no observed views after sampling, one view is randomly restored to ensure that at least one modality remains available. Thus, every sample satisfies ∣Ωi∣≥1|\Omega_i|\geq1, where Ωi\Omega_i denotes its observed-view set. This condition ensures that each sample can provide at least one source representation for subsequent learning and retrieval.

The missing mask is independently generated for every run and remains fixed throughout that run. The same mask is used consistently during representation learning, graph construction, target-aware retrieval, pseudo-missing completion training, and final clustering evaluation. In particular, the mask is not regenerated between training epochs or between the retrieval and clustering stages. This protocol ensures that the reported results correspond to a well-defined incomplete-view realization for each independent run.

The five-run protocol is applied separately to every missing rate. In the main experiments, the seven datasets are evaluated at missing rates 0.10.1, 0.30.3, 0.50.5, and 0.70.7. Appendix C.2 additionally evaluates missing rates 0.20.2, 0.40.4, and 0.60.6, using the same five-run protocol. Therefore, the experiments cover seven missing-rate settings in total. For every dataset-rate combination, five independently generated masks and five independently trained models are evaluated.

For each setting, the reported ACC, NMI, and ARI values are calculated as the mean over the five runs, and the corresponding standard deviations are also reported. Table 1 follows this convention for the main comparison, while Table 7 applies the same reporting protocol to the additional missing rates. The ablation and runtime experiments likewise use five independent runs. Therefore, the reported variation reflects differences in random initialization, missing-mask generation, and stochastic optimization across independent experiments.

The independent runs use the same experimental configuration and hyperparameters to ensure a controlled comparison. However, each run has its own random seed and independently sampled observation mask. The missing mask remains unchanged within an individual run, so the standard deviation does not arise from repeatedly changing the observation pattern during a single training process. Instead, it measures the variation across independent realizations of the incomplete multi-view learning problem.

The protocol is specified in Section 4.1 and Appendix B.2. Section 4.1 states the five-run evaluation setting, while Appendix B.2 describes the missing-mask sampling procedure and the condition that every sample retains at least one observed view. The supplementary experiments in Appendix C.2 follow the same protocol, ensuring consistency between the main and additional results.

For completion training, pseudo-missing targets are selected only from originally observed entries. The selected target view is temporarily hidden when the retrieval query and candidate set are constructed, while its original latent representation serves as a self-supervised training target. At inference, pseudo-missing masking is disabled, and only genuinely missing sample-view pairs are completed. Ground-truth cluster labels are not used to generate missing masks, train the completion module, construct retrieval queries, or obtain the final cluster assignments.

Accordingly, the number of independent repeated experiments is five for every missing rate on every dataset. All reported means and standard deviations are calculated from these five independent runs, and the same protocol is maintained throughout the main experiments, supplementary comparisons, and ablation studies.

AuthorsOpenReview ↗

Q2. Parameter Sensitivity and Stable Clustering Performance

Q2: We thank the reviewer for examining the parameter sensitivity results and the opportunity to explain the stable performance. Figure 7 evaluates the influence of two objective weights, λs\lambda_s and λg\lambda_g. The broad high-performing regions show that TIGER does not rely on a narrowly selected combination.

The overall objective is L=Lrec+λsLs+λgLgL=L_{rec}+\lambda_s L_s+\lambda_g L_g, where LrecL_{rec} denotes reconstruction, LsL_s denotes graph-structure decoupling, and LgL_g denotes graph-intervention completion. The weights λs\lambda_s and λg\lambda_g balance two complementary objectives. The first aligns neighborhood distributions across jointly observed views and separates stable from private information. The second improves target-view recovery under private-edge interventions and regulates uncertain corrections through retrieval confidence.

Figure 7 in Appendix C.5 presents four datasets, HW, MNIST4, MNIST-Fea, and Synthetic3D, under five missing rates: r=0.1,0.3,0.4,0.5,0.7r=0.1,0.3,0.4,0.5,0.7. This produces twenty parameter-response surfaces, each reporting ACC over the tested combinations of λs\lambda_s and λg\lambda_g. The main manuscript also reports results at a missing rate of 0.1. One parameter configuration is shared across datasets.

The two weights do not act as independent switches that determine separate cluster labels. Instead, they scale learning signals within the same representation pipeline. When either coefficient changes moderately, the encoder can still preserve the dominant neighborhood organization while adjusting the contribution of the corresponding objective. As a result, the final partition can remain stable even though the optimization trajectory and latent vectors are not exactly the same. A larger change can move the model beyond this stable region and alter sample assignments, which explains the local variations visible in the response surfaces.

The broad high-performing regions have a direct interpretation. The objectives preserve relational information and improve missing-view recovery. Moderate changes in their weights can alter the relative optimization emphasis without necessarily changing final cluster assignments. Consequently, ACC may remain similar across neighboring configurations even when learned representations and internal objective values differ.

This behavior is reasonable because ACC is computed from discrete clustering assignments, not directly from continuous latent representations. K-means converts embeddings into cluster labels. A parameter change may modify local distances or embedding geometry without changing the final assignment of individual samples. In that case, ACC remains unchanged. Small changes in correct assignments may produce limited percentage differences. Therefore, a relatively flat ACC surface does not imply that the parameters have no influence or that the learned representations are identical.

The surfaces do not indicate that every parameter combination performs equally well. Some regions show visible changes, particularly near the boundaries of high-performing areas. These variations confirm that the coefficients affect the balance between relational alignment and completion. The broad high-performing regions instead indicate that TIGER does not require extremely precise coefficient tuning to maintain good clustering quality. This is useful across datasets with different feature distributions and missingness patterns.

The experiments use the shared configuration λs=0.45\lambda_s=0.45 and λg=1.15\lambda_g=1.15. This avoids dataset-specific choices and provides consistent benchmark comparisons. The sensitivity surfaces show that this configuration lies within a useful high-performing region rather than depending on an isolated optimum. The analysis tests robustness to parameter perturbations.

Appendix C.6 provides complementary convergence evidence. Figure 8 reports training trajectories for Synthetic3D, HW, MNIST4, and ALOI-100 at the same five missing rates. The curves approach stable regions with smaller later fluctuations. Despite lower performance under severe missingness, the trajectories remain stable without persistent oscillations or divergence.

Appendix C.5 also examines neighborhood size kk and the Smooth L1 transition parameter δ\delta in Figure 6. The results show relatively stable clustering performance across broad tested ranges at missing rates 0.1 and 0.3. Thus, the analysis also covers neighborhood retrieval and reconstruction parameters.

The revised manuscript presents these results in Figures 6–8 and Appendices C.5–C.6. Together, the experiments show that TIGER maintains strong clustering performance over a useful range of parameter configurations and missing rates. Stable ACC regions are consistent with discrete clustering evaluation and demonstrate practical robustness. Local variations remain visible, indicating that parameters matter but the selected configuration is not excessively sensitive within the tested range.

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

This paper proposes TIGER, a target-aware retrieval framework for incomplete multi-view clustering. It separates stable and view-specific representations to establish transferable structural relationships, then retrieves target-view anchors for missing-view recovery. A confidence-guided graph intervention mechanism further improves completion robustness through stochastic perturbation and residual correction. Experiments on seven benchmark datasets demonstrate consistently strong clustering performance across different missing rates, supported by ablation and sensitivity analyses.

Critical strengths
  1. The paper addresses important challenges in incomplete multi-view clustering, particularly unreliable structural transfer and missing-view recovery. The proposed target-aware formulation provides a clear and meaningful research motivation.
  2. The integration of graph-structure decoupling, target-view anchor retrieval, and confidence-guided completion offers a coherent framework for exploiting transferable structural information.
  3. Experiments on seven benchmark datasets against recent competitive baselines demonstrate consistently strong clustering performance. The reported improvements under severe missingness further support the practical effectiveness of the approach.
Critical weaknesses
  1. Several components resemble established disentangled representation learning, anchor retrieval, and consistency regularization techniques, while their distinct contributions are insufficiently demonstrated.
  2. The graph-structure decoupling objective aligns neighborhood distributions but does not explicitly guarantee cross-view feature-space alignment. This raises concerns about the validity of directly aggregating stable representations for target-aware retrieval.
  3. The ablation studies mainly examine combinations of training losses rather than individual mechanisms. Additional experiments are needed to isolate the effects of target-aware retrieval, confidence gating, and graph intervention.
Questions for authors
  1. Could the authors clarify how stable representations from different views are aligned in a shared coordinate space, given that neighborhood-distribution consistency alone does not guarantee feature-level alignment?
  2. Could the authors provide separate ablations for target-aware anchor retrieval, confidence gating, and graph intervention to establish their individual contributions?
  3. Could the authors evaluate completion accuracy and retrieval confidence calibration, and report computational costs and performance under non-random missingness patterns?
7 comments
AuthorsOpenReview ↗

W1. Novelty and Individual Contributions of the Proposed Components

W1: We thank the reviewer for the careful assessment of the relationship between TIGER and established disentangled representation learning, anchor retrieval, and consistency regularization. TIGER coordinates these ideas for incomplete multi-view clustering: transferable neighborhood structure guides target-specific retrieval, and retrieval confidence controls view-specific evidence during completion.

Equation (3) decomposes each view-specific latent representation into stable and private factors, siv=Fsv(ziv)s_i^v=F_s^v(z_i^v) and piv=Fpv(ziv)p_i^v=F_p^v(z_i^v). Stable factors preserve relational information that can be compared across views, while private factors retain view-dependent variation. This factorization separates transferable relations from target-specific evidence within the recovery pipeline.

Equations (4) and (5) define the graph-structure decoupling objective. For a pair of views, neighborhood distributions are constructed over jointly observed samples, using the same sample indices in both views. The term JSD⁡(πivu,πiuv)\operatorname{JSD}(\pi_i^{vu},\pi_i^{uv}) encourages consistency between their relative neighborhood structures, while ∣cos⁡(siv,piv)∣|\operatorname{cos}(s_i^v,p_i^v)| discourages redundant information between stable and private factors. It transfers local relations without requiring point-wise equality between feature vectors.

The aligned stable factors then provide the retrieval descriptor. When view vv is missing for sample ii, Eq. (6) aggregates stable factors from the available views and normalizes the result to form query qivq_i^v. The query is a retrieval descriptor, not an imputed target feature. Equation (7) searches only the observed target-view bank and selects the top-kk candidates by cosine similarity. The source views determine the query, but candidate samples and the codes used to construct anchors come from the target view. Source-view neighbors are therefore not substituted for target-view observations.

Equations (8) and (9) construct two complementary anchors from the selected target-view candidates. Stable weights αijv\alpha_{ij}^v produce anchor rivr_i^v, which estimates the principal target-view location. Private weights βijv\beta_{ij}^v produce anchor uivu_i^v, which retains view-specific variations that stable aggregation alone may smooth.

Equation (10) measures the concentration of stable retrieval weights through normalized entropy. A concentrated distribution gives higher confidence, whereas a diffuse distribution gives lower confidence. This score is a retrieval-concentration measure, not a calibrated probability of reconstruction correctness. Equation (11) applies two independent stochastic interventions to private candidate edges while keeping the stable anchor fixed. The remaining private weights are renormalized, with at least one candidate retained in each pass. The interventions test dependence on particular private relations. Equation (12) uses confidence to gate the private residual, yiv,t=riv+civGv([riv∥uiv,t])y_i^{v,t}=r_i^v+c_i^vG^v([r_i^v\Vert u_i^{v,t}]). The stable anchor supplies the target reference, while confidence controls the private correction. Averaging the two predictions yields the completed representation.

The individual contributions are examined in the revised component ablation in Table 3. It removes graph intervention, target-aware retrieval, and confidence gating separately, rather than evaluating only combinations of training losses. At a missing rate of 0.1, the full model obtains the highest mean ACC, NMI, and ARI on ALOI-100, MNIST-Fea, and MNIST4. For example, on ALOI-100, removing retrieval reduces ACC from 79.55 to 77.44, while removing confidence reduces it to 76.96. Removing graph intervention reduces it to 76.03. On MNIST4, the full model reaches 91.03 ACC, compared with 88.99 without graph intervention, 87.99 without retrieval, and 86.99 without confidence. The corresponding NMI and ARI results show the same overall advantage of the complete design.

The full model leads in all nine metric comparisons, with means and standard deviations over five independent runs. The results also show that the modules are not interchangeable: retrieval supplies target-view candidates, confidence regulates their private contribution, and graph intervention tests the stability of that contribution under perturbed relations. This supports their complementary roles across all three datasets.

Table 2 separately evaluates objective combinations, while Appendix C.4 extends the objective ablation to Synthetic3D, HW, and CUB. Table 2 assesses training objectives, whereas Table 3 isolates recovery mechanisms. Together, they show that the complete framework benefits from relational alignment, target-specific evidence, confidence-controlled correction, and robustness to private-edge perturbations. TIGER's contribution is therefore the integrated mechanism and its demonstrated component-level effect, rather than a claim of inventing disentanglement, retrieval, or consistency regularization individually.

AuthorsOpenReview ↗

W2. Neighborhood Alignment and Cross-View Representation Compatibility

W2: We are grateful for the opportunity to clarify how stable representations from different views are used and what the graph-structure objective actually guarantees. TIGER does not claim that neighborhood-distribution consistency establishes exact feature-level alignment or makes the coordinates of all views identical. Each view retains its own encoder and factorization functions. The common latent dimension provides a compatible representation size for subsequent operations, but equal dimensionality alone is not interpreted as semantic coordinate alignment.

Equations (4) and (5) compare neighborhood distributions constructed over jointly observed sample indices. For views vv and uu, the set IvuI_{vu} contains samples observed in both views. Similarities among stable factors are normalized to obtain πivu\pi_i^{vu} and πiuv\pi_i^{uv}. The Jensen-Shannon term compares the relative probability assigned to the same sample neighbors. It encourages the two views to preserve compatible local relations, while the stable-private term reduces redundancy within each view. Neither term imposes siv=sius_i^v=s_i^u, nor does it require every coordinate in one view to correspond to a coordinate in another view.

The shared dimension d=64d=64 is an interface for the stable factors and subsequent retrieval operations, not evidence that the original view-specific feature spaces have become identical. The encoders remain view-specific, and the input dimensions dvd_v may differ. Their outputs can therefore have a common length while retaining different coordinate semantics. The method uses relational comparability, not an assumption of fully aligned feature coordinates.

This distinction matters for Eq. (6). When target view vv is missing, the stable factors from available views are aggregated and normalized to form query qivq_i^v. This query is a cross-view retrieval descriptor that summarizes the available relational evidence. It is not treated as a completed target vector, and its coordinates are not directly substituted for target-view features. The query only determines which observed target samples are relevant candidates.

Equation (7) searches the observed target-view bank IvI_v using the query and the stable factors of target-view candidates. Thus, the comparison is performed within the target view's own representation branch. The candidate codes used in Eqs. (8) and (9) also come from that target view. The stable anchor rivr_i^v is formed by weighting observed target stable codes, and the private anchor uivu_i^v is formed from observed target private codes. Consequently, the actual reconstruction evidence is grounded in target-view observations rather than in a direct sum of source-view feature vectors.

The source-view aggregation and target-view retrieval serve different purposes. The former creates a query that can exploit all available views; the latter validates and selects candidates according to target-view geometry. This two-stage design reduces the risk of treating a neighbor in one source view as an appropriate target-view neighbor. The retrieved target codes determine both anchors, so the final completion remains specific to the missing view.

Equation (10) evaluates the concentration of stable retrieval weights. Higher concentration produces higher confidence, while a diffuse candidate distribution produces lower confidence. This confidence controls the private residual in Eq. (12), rather than changing the stable target anchor. Equations (11) and (12) further test the private evidence under two independent edge interventions. The prediction is yiv,t=riv+civGv([riv∥uiv,t])y_i^{v,t}=r_i^v+c_i^vG^v([r_i^v\Vert u_i^{v,t}]). The stable anchor provides the target-view reference, and confidence regulates the amount of private correction. These steps do not require source and target coordinates to be identical.

The distinction is also reflected in the experimental design. The method is evaluated through clustering performance under incomplete observations, and the retrieval candidates are restricted to observed target samples. The ablations in Table 3 show the effect of removing target-aware retrieval, confidence gating, and graph intervention separately. The complete model achieves the best mean ACC, NMI, and ARI in all nine comparisons on ALOI-100, MNIST-Fea, and MNIST4. These results support the role of target-specific retrieval and confidence-controlled completion, rather than an unverified assumption of coordinate-wise alignment.

Therefore, the stable representations are made comparable through shared sample-neighborhood relations and a common output dimension, not through an explicit coordinate-matching loss. The query is used to retrieve target-view evidence, and target-view observations construct the anchors and completed representation. This is the intended meaning of cross-view relational transfer in TIGER. The method preserves view-specific feature semantics while exploiting common relational structure to guide recovery.

AuthorsOpenReview ↗

W3. Individual Contributions of the Proposed Mechanisms

W3: We thank the reviewer for highlighting the importance of isolating the individual mechanisms in addition to evaluating combinations of training objectives. The revised analysis separates objective-level and mechanism-level ablations. Table 2 examines combinations of optimization losses, whereas Table 3 separately removes target-aware retrieval, confidence gating, and graph intervention. The former evaluates training signals; the latter isolates recovery mechanisms.

The ablation uses ALOI-100, MNIST-Fea, and MNIST4 at a missing rate of 0.1. Each configuration is evaluated over five independent runs, with mean and standard deviation reported. "No Graph" removes the private-edge graph intervention, "No Retrieval" removes target-aware anchor retrieval, and "No Confidence" removes confidence-based gating of the private correction. The Full configuration retains all three mechanisms.

ALOI-100

VariantACCNMIARI
No Graph76.03 ±\pm 2.5983.98 ±\pm 1.4065.99 ±\pm 2.67
No Retrieval77.44 ±\pm 0.8184.44 ±\pm 0.0666.32 ±\pm 1.34
No Confidence76.96 ±\pm 0.6683.15 ±\pm 0.2365.19 ±\pm 1.47
Full79.55 ±\pm 0.6885.28 ±\pm 0.6768.61 ±\pm 2.09

MNIST-Fea

VariantACCNMIARI
No Graph96.48 ±\pm 0.0093.68 ±\pm 0.0294.68 ±\pm 0.01
No Retrieval97.48 ±\pm 0.0094.68 ±\pm 0.0295.68 ±\pm 0.01
No Confidence98.20 ±\pm 0.2894.06 ±\pm 0.6095.08 ±\pm 0.60
Full98.48 ±\pm 0.0095.68 ±\pm 0.0296.68 ±\pm 0.01

MNIST4

VariantACCNMIARI
No Graph88.99 ±\pm 0.8674.60 ±\pm 2.3073.63 ±\pm 1.95
No Retrieval87.99 ±\pm 0.8674.58 ±\pm 2.2971.62 ±\pm 1.93
No Confidence86.99 ±\pm 0.8672.58 ±\pm 2.2873.63 ±\pm 1.95
Full91.03 ±\pm 0.9077.68 ±\pm 2.3478.71 ±\pm 2.03

Full obtains the highest mean ACC, NMI, and ARI in all nine comparisons. On ALOI-100, removing graph intervention, retrieval, and confidence gating reduces ACC from 79.55 to 76.03, 77.44, and 76.96, respectively. On MNIST-Fea, the corresponding values decrease from 98.48 to 96.48, 97.48, and 98.20. On MNIST4, ACC decreases from 91.03 to 88.99, 87.99, and 86.99. NMI and ARI show the same overall pattern, indicating improved assignments and cluster-structure agreement.

The three mechanisms have distinct functions and are not interchangeable. Target-aware retrieval uses the stable query to select candidates exclusively from the observed target-view bank. It provides target-specific evidence and constructs stable and private anchors from actual target-view representations. Confidence gating evaluates the concentration of stable retrieval weights and controls the contribution of the private residual. It prevents diffuse retrieval evidence from producing an equally strong private correction. Graph intervention perturbs private candidate edges while retaining the stable anchor, testing whether completion depends excessively on particular view-specific relations. Removing each mechanism therefore removes a different capability from the recovery process.

On ALOI-100, the Full model improves ACC, NMI, and ARI over every individual ablation. On MNIST-Fea, where ACC is already high for all variants, the improvements in NMI and ARI provide additional evidence of better cluster organization. On MNIST4, the full configuration produces clear gains in all metrics, including ARI, which evaluates pairwise agreement between cluster assignments. The contribution is not confined to one metric or dataset.

Table 2 complements this component analysis by evaluating objective combinations, including reconstruction, graph-structure decoupling, and graph-intervention completion. Appendix C.4 further extends objective-level comparisons to Synthetic3D, HW, and CUB. Table 2 therefore studies which training signals are useful, while Table 3 studies whether the individual retrieval and completion mechanisms contribute to the final result. This separation distinguishes loss contributions from inference-time recovery effects.

The gains of the Full model over the strongest ablated result are 2.11, 0.84, and 2.29 percentage points for ACC, NMI, and ARI on ALOI-100; 0.28, 1.00, and 1.00 points on MNIST-Fea; and 2.04, 3.08, and 5.08 points on MNIST4. The full configuration therefore leads in every reported metric comparison. Five-run standard deviations show variability beyond single-run values.

These results provide direct component-level evidence for target-aware retrieval, confidence-controlled correction, and private-edge robustness. Together, the ablations show that TIGER benefits from complementary recovery mechanisms, not merely combined training losses. The separate rows make each component's contribution directly inspectable. The revised Table 3 and the corresponding discussion report these separate effects.

Reviewer_PDWSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing IMVC methods face three key challenges: 1) feature reconstruction or global information sharing may ignore the distribution characteristics of target views; 2) structural relationships among samples and complementary cross-view information are insufficiently explored during missing-view recovery; 3) there is a lack of explicit mechanisms to evaluate the reliability of recovered information, so uncertain completion results negatively affect representation fusion and clustering optimization. To tackle these issues, the authors proposed TIGER, which first learned view-specific latent representations and separates transferable from view-dependent information to build more reliable cross-view relationships. Besides this, an intervention-based completion mechanism is further introduced to identify and reduce unreliable local correlations, enabling robust representation recovery under severe missing conditions.

Critical strengths
  1. It proposed a target-aware graph retrieval framework for IMVC, which uses transferable structural evidence to locate anchors in the target-view bank and recasts missing-view recovery as target-grounded retrieval instead of source-to-target feature reconstruction.
  2. It introduced confidence-guided stochastic graph intervention and residual completion to challenge fragile private relations and to adaptively control correction based on retrieval reliability, thereby reducing the impact of uncertain completion on fusion and clustering.
Critical weaknesses
  1. The paper said it is "the first to introduce graph-structure decoupling into incomplete multi-view clustering". Firstly, disentanglement learning of specific feature is not a new proposed method. Secondly, the superiority of the proposed graph-structure decoupling approach should be further proved.
  2. The proposed model is very complex, and the corresponding time complexity analysis is not sufficient. In the experiment, it is necessary to supplement a large-scale database to verify that the running time is acceptable.
  3. Many the hightest experimental results are very close to the suboptimal ones, such that the contributions of the proposed method is suspicious.
Questions for authors

Referring to the weaknesses.

4 comments
AuthorsOpenReview ↗

Novelty and Evidence for Graph-Structure Decoupling

W1: We thank the reviewer for questioning the novelty claim and the evidence for graph-structure decoupling. The contribution lies in transferable sample relations for incomplete multi-view clustering, not in claiming that separating shared and view-specific features is new. TIGER connects stable-private factorization to neighborhood alignment, target-view retrieval, and confidence-controlled completion. Feature separation alone does not determine which relations can transfer across views or guide recovery of a missing target.

For an observed sample, the encoder maps latent vector zivz_i^v to stable factor sivs_i^v and private factor pivp_i^v. The stable factor captures transferable local structure, while the private factor retains view-dependent information. The factors need not share identical coordinates across views. Equation (4) instead constructs neighborhood distributions over jointly observed samples Ivu=Iv∩IuI_{vu}=I_v\cap I_u. Pairwise cosine similarities between stable factors become normalized probabilities πijvu\pi_{ij}^{vu}, using the same sample indices in both views. The comparison remains meaningful across heterogeneous feature spaces.

Equation (5) combines two constraints. The Jensen-Shannon term JSD⁡(πivu,πiuv)\operatorname{JSD}(\pi_i^{vu},\pi_i^{uv}) aligns relative neighborhood distributions, while ∣cos⁡(siv,piv)∣|\cos(s_i^v,p_i^v)| reduces redundancy between stable and private factors. TIGER therefore transfers local topology without imposing point-wise equality such as siv=sius_i^v=s_i^u. This is the graph-level distinction from feature disentanglement: the stable branch preserves cross-view relational evidence, while the private branch retains view-specific variation. Only jointly observed samples enter graph construction.

The aligned stable factors guide target-view recovery. When view vv is missing for sample ii, Eq. (6) aggregates stable factors from its available views and normalizes them to form query qivq_i^v. The query is a retrieval descriptor, not an imputed target feature. Equation (7) searches only the observed target-view bank IvI_v for its top-kk candidates. Source views determine the query, while candidates and target codes come from the target view.

Equations (8) and (9) aggregate candidates into two anchors. Stable weights αijv\alpha_{ij}^v form anchor rivr_i^v, which estimates target-view location; private weights βijv\beta_{ij}^v form anchor uivu_i^v, which retains view-specific variation. Equation (10) computes normalized-entropy confidence from the stable weights. Concentrated weights give higher confidence, while diffuse weights give lower confidence. The score measures retrieval concentration, not calibrated correctness. Equation (11) applies two independent stochastic interventions to private candidate edges while keeping the stable anchor fixed. Remaining private weights are renormalized, with at least one candidate retained per pass. Equation (12) gates the private residual: yiv,t=riv+civGv([riv∥uiv,t])y_i^{v,t}=r_i^v+c_i^vG^v([r_i^v\Vert u_i^{v,t}]). The stable anchor supplies the target reference, confidence controls correction, and intervention tests dependence on fragile private relations. Predictions are averaged.

Table 3 isolates graph intervention, target-aware retrieval, and confidence gating on three datasets at a missing rate of 0.1. The entries below are mean and standard deviation over five runs; each strongest ablation is selected separately for each metric.

DatasetMetricStrongest ablationFull TIGERGain
ALOI-100ACC77.44 ±\pm 0.8179.55 ±\pm 0.682.11
ALOI-100NMI84.44 ±\pm 0.0685.28 ±\pm 0.670.84
ALOI-100ARI66.32 ±\pm 1.3468.61 ±\pm 2.092.29
MNIST-FeaACC98.20 ±\pm 0.2898.48 ±\pm 0.000.28
MNIST-FeaNMI94.68 ±\pm 0.0295.68 ±\pm 0.021.00
MNIST-FeaARI95.68 ±\pm 0.0196.68 ±\pm 0.011.00
MNIST4ACC88.99 ±\pm 0.8691.03 ±\pm 0.902.04
MNIST4NMI74.60 ±\pm 2.3077.68 ±\pm 2.343.08
MNIST4ARI73.63 ±\pm 1.9578.71 ±\pm 2.035.08

The full model leads in all nine comparisons. Table 2 separately evaluates reconstruction, graph-structure decoupling, and graph-intervention completion; the full objective obtains the best ACC, NMI, and ARI on MNIST4, ALOI-100, and MNIST-Fea. Appendix C.4 extends objective-level ablation to Synthetic3D, HW, and CUB, where the full objective also leads. The gains therefore reflect the combined roles of relational alignment, target-specific retrieval, and confidence-guided intervention.

TIGER's novelty is thus the coordinated use of stable-private representations to align transferable neighborhood structure and guide target-view retrieval, followed by confidence-gated completion under private-edge interventions. The claim concerns this integrated mechanism, not feature disentanglement in general. Sections 3.2–3.4, Tables 2–3, and Appendix C.4 provide the method details and component-level evidence.

AuthorsOpenReview ↗

W2. Computational Complexity and Runtime Evaluation

W2: We appreciate the reviewer’s focus on computational cost and the request for a clearer runtime account. The revised manuscript provides a component-wise complexity analysis in Appendix B.4, covering representation learning, graph construction, retrieval, completion, inference, and storage. Tables 4 and 5 additionally report measured end-to-end runtime on ALOI-100 and MNIST-Fea.

Let IvI_v denote the observed sample set in view vv, and Ivu=Iv∩IuI_{vu}=I_v\cap I_u the jointly observed set for a view pair. The input, hidden, and latent dimensions are dvd_v, hh, and dd. Let tvt_v count pseudo-missing training targets in view vv, and mvm_v count actually missing sample-view pairs at inference. For readability, the complexity terms are first summarized by component symbols.

ComponentComplexity term
Representation learningCrepC_{\mathrm{rep}}
Dense graph constructionCgraphC_{\mathrm{graph}}
Retrieval and completionCret(n)C_{\mathrm{ret}}(n)
Training per epochO(Crep+Cgraph+Cret(t))O(C_{\mathrm{rep}}+C_{\mathrm{graph}}+C_{\mathrm{ret}}(t))
Inference, including clusteringO(Crep+Cret(m)+TKNKd)O(C_{\mathrm{rep}}+C_{\mathrm{ret}}(m)+T_KNKd)
SpaceCspaceC_{\mathrm{space}}

Representation learning includes the view-specific encoder and decoder: Crep=O(∑v=1V∣Iv∣(dvh+h2+hd+d2))C_{\mathrm{rep}}=O(\sum_{v=1}^{V} |I_v|(d_vh+h^2+hd+d^2)). Dense graph construction compares stable representations among jointly observed samples: Cgraph=O(∑(v,u)∈P∣Ivu∣2d)C_{\mathrm{graph}}=O(\sum_{(v,u)\in P}|I_{vu}|^2d). For nvn_v targets in each view, retrieval, candidate aggregation, and the two private-edge intervention passes have cost: Cret(n)=O(∑v=1Vnv(∣Iv∣d+kd+hd+h2))C_{\mathrm{ret}}(n)=O(\sum_{v=1}^{V}n_v(|I_v|d+kd+hd+h^2)). Training uses nv=tvn_v=t_v, while inference uses nv=mvn_v=m_v; thus, only actually missing entries are retrieved at inference. The two private-edge interventions reuse the stable neighborhood and add a constant factor.

The inference expression includes K-means cost TKNKdT_KNKd, where TKT_K, NN, and KK are its iteration count, sample number, and cluster number. The space term is: Cspace=O(∑v=1V∣Iv∣d+∑(v,u)∈P∣Ivu∣2+∑v∑i:Miv=0(k+d))C_{\mathrm{space}}=O(\sum_{v=1}^{V}|I_v|d+\sum_{(v,u)\in P}|I_{vu}|^2+\sum_{v}\sum_{i:M_i^v=0}(k+d)). In the exact dense implementation, the graph term is quadratic in the number of jointly observed samples and is the principal scaling bottleneck. With fixed kk, target-bank scanning is linear in observed target samples per missing entry. Sparse storage replaces the quadratic term with the retained-edge count.

The representation term covers per-view encoder and decoder passes and the stable and private branches. The graph term evaluates stable-factor similarities for each valid view pair over jointly observed indices. The retrieval term includes query-to-bank similarity, top-kk selection, anchor aggregation, and two intervention-based predictions. Its cost depends on the number of missing targets, not the full sample count. Therefore, fewer missing entries reduce the retrieval-and-completion component at inference.

All costs are stated in asymptotic form. These expressions describe the dense implementation used in the experiments. They distinguish the quadratic graph-construction cost from the per-missing-target bank scan, making the principal source of scaling cost explicit.

For empirical runtime, Tables 4 and 5 compare TIGER with GAVIM at missing rates 0.10.1, 0.30.3, 0.50.5, and 0.70.7. Each setting is repeated five times; mean runtime is reported in seconds.

DatasetMissing rateGAVIM (s)TIGER (s)Speedup
ALOI-1000.12281.45137.0216.65x
ALOI-1000.33925.65381.3510.29x
ALOI-1000.52114.74116.8618.10x
ALOI-1000.71750.1586.7420.18x
MNIST-Fea0.13145.17566.675.55x
MNIST-Fea0.35013.61801.606.25x
MNIST-Fea0.53596.27428.978.38x
MNIST-Fea0.72053.75223.969.17x

Across all eight settings, TIGER is faster by 5.555.55 to 20.1820.18 times and also obtains higher ACC than GAVIM. On ALOI-100, the ACC gains are 3.883.88, 14.4914.49, 23.2223.22, and 34.5934.59 percentage points at the four respective missing rates. On MNIST-Fea, the gains are 0.610.61, 2.272.27, 4.414.41, and 6.716.71 points. Thus, the runtime advantage does not come at the expense of clustering accuracy.

The two benchmarks provide runtime evidence, not a large-scale evaluation. The study focuses on reliable target-view recovery and clustering under incomplete observations. The analysis identifies dense pairwise graph construction as the main component to address when scaling to larger sample counts, while retrieval has a separate linear scan cost per missing target. Each sample retains at least one observed view, and each target bank contains at least k=36k=36 observed samples. Appendix B.4 gives the derivation; Tables 4 and 5 report runtime.

AuthorsOpenReview ↗

W3. Interpretation of Close Best and Second-Best Results

W3: We thank the reviewer for examining the gaps between the best and second-best results and for asking whether the improvements translate into a substantive contribution. An isolated small margin is not the sole basis for the claim. TIGER's evidence is instead based on repeated-run means across the complete benchmark matrix, additional intermediate missing rates, and controlled ablations that identify where the improvements come from.

Table 1 reports mean and standard deviation over five independent runs for seven datasets, four missing rates, and three metrics. This gives 84 dataset-rate-metric comparisons. TIGER achieves the highest mean in all 84 settings. The size of the margin depends on dataset difficulty and missingness; some settings are close, while challenging settings show larger gains. The following examples report the strongest competing value separately for each metric, rather than using one baseline for all three metrics.

DatasetMissing rateMetricStrongest baselineTIGERGain
Synthetic3D0.1ACC96.1496.500.36
Synthetic3D0.1NMI85.1285.580.46
Synthetic3D0.1ARI89.0489.860.82
HW0.7ACC87.4588.801.35
HW0.7NMI78.6379.100.47
HW0.7ARI77.8679.251.39
ALOI-1000.7ACC48.1771.8823.71
ALOI-1000.7NMI62.2476.0813.84
ALOI-1000.7ARI50.4356.706.27

These examples show why the results should be considered across datasets and metrics. On Synthetic3D, the margins are modest but TIGER remains best on all three metrics. On HW, the gains are consistent across ACC, NMI, and ARI under severe missingness. On ALOI-100, where many categories and visual variation make target-view recovery more difficult, the margins become substantially larger. The method is not supported by selecting only large-margin cases; the full table reports every dataset and missing-rate combination.

Appendix C.2 adds missing rates 0.2, 0.4, and 0.6, producing 63 additional comparisons. TIGER again obtains the highest mean in all 63 settings. For example, on ALOI-100 at a missing rate of 0.6, the gains over the strongest competing results are 23.02 points in ACC, 14.07 points in NMI, and 10.32 points in ARI. Together, the main and additional tables contain 147 comparisons, and the same five-run reporting protocol is used throughout. The standard deviations make run-to-run variation visible; the results are not presented as significance tests, and a small numerical gap is not treated as proof of statistical significance by itself. ACC, NMI, and ARI are computed from the final discrete partition. Changes in embedding geometry or retrieval confidence can improve local organization without changing cluster assignments, leaving the final score close. The ablations therefore test component contributions beyond the ranking of the final score.

The component ablations provide evidence beyond the ranking of competing methods. Table 3 removes target-aware retrieval, confidence gating, and graph intervention separately. The full model achieves the best mean in all nine metric comparisons on ALOI-100, MNIST-Fea, and MNIST4. The gains over the strongest ablated result are summarized below.

DatasetACC gainNMI gainARI gain
ALOI-1002.110.842.29
MNIST-Fea0.281.001.00
MNIST42.043.085.08

Table 2 separately evaluates the objective combinations. The complete objective achieves the best ACC, NMI, and ARI on MNIST4, ALOI-100, and MNIST-Fea. Appendix C.4 extends this analysis to Synthetic3D, HW, and CUB, where the complete objective also leads across the reported metrics. These ablations show that the full method's performance is not merely a consequence of adding parameters: target-aware retrieval supplies target-view evidence, confidence gating controls uncertain private corrections, and graph intervention improves robustness to view-specific edges. Removing any one of these mechanisms reduces at least part of the measured performance.

The practical contribution is also reflected in runtime. Tables 4 and 5 show that TIGER has lower total runtime than GAVIM in all eight dataset-rate settings, with speedups from 5.55x to 20.18x, while obtaining higher ACC in every setting. Thus, the evidence is not limited to close comparisons in a few cells. It combines consistent best mean results across 147 settings, five-run variability, objective-level and component-level ablations, and a favorable accuracy-efficiency trade-off. We therefore interpret the small margins as evidence of competitive performance on relatively easier settings, while the ablations and more challenging missing-view settings establish the contribution of the proposed mechanisms. The full results and standard deviations remain available in Tables 1–5 and 7–8 for every reported condition.

AuthorsOpenReview ↗

Clarifying the Novelty and Practical Effectiveness of TIGER

Dear Reviewer,

We sincerely appreciate your careful evaluation and constructive feedback. Your comments have encouraged us to further clarify the methodological contributions of TIGER and strengthen the experimental evidence supporting its effectiveness.

In response to your concerns, we have made several important revisions:

  • We clarified the distinction between conventional feature disentanglement and our graph-structure decoupling mechanism. Specifically, stable-private factorization serves as a representation foundation, while our contribution lies in aligning cross-view neighborhood distributions and exploiting transferable structural relations for target-aware retrieval.
  • We expanded the computational complexity analysis in Appendix B.4 to provide a more comprehensive discussion of the computational and memory requirements of TIGER.
  • We added runtime comparisons against GAVIM on ALOI-100 and MNIST-Fea across four missing rates. TIGER consistently achieves lower runtime, with speedups ranging from 5.55× to 20.18×, while maintaining superior clustering performance.
  • We introduced additional component-level ablation experiments to distinguish the contributions of target-aware retrieval, confidence gating, and graph intervention. We also explicitly reported five independent runs for each missing-rate setting, together with the corresponding mean and standard deviation.

We believe these revisions provide clearer evidence of the methodological contributions, computational efficiency, and experimental reliability of TIGER.

We sincerely hope that these improvements adequately address your concerns and clarify the effectiveness of the proposed approach.

Thank you again for your valuable feedback and consideration.

Sincerely,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

Dear PCs, SACs, ACs, and Reviewers,

We sincerely thank all reviewers for their constructive feedback and valuable suggestions, which have helped us improve the clarity, technical rigor, and completeness of our manuscript.

In response to the reviewers' concerns, we have made substantial efforts to refine the methodological descriptions, conduct additional experiments, and strengthen the empirical analyses. The major updates are summarized below:

  • We clarified the methodological foundations of TIGER, including the distinction between transferable graph topology and view-specific variations, the role of target-view retrieval, and the relationship between retrieval confidence and stochastic graph intervention.
  • We further clarified the applicability and assumptions of target-aware retrieval, explaining how partially observed target-view banks support anchor selection without accessing missing target features.
  • We strengthened the experimental protocol by explicitly conducting five independent runs with different random seeds for each missing-rate setting and reporting the corresponding mean and standard deviation.
  • We added component-level ablation experiments on ALOI-100, MNIST-Fea, and MNIST4 to independently evaluate the contributions of target-aware retrieval, graph intervention, and confidence gating.
  • We introduced additional runtime comparisons against GAVIM on ALOI-100 and MNIST-Fea across four missing rates, demonstrating that TIGER achieves a favorable balance between clustering performance and computational efficiency, with speedups ranging from 5.55× to 20.18×.
  • We further refined the analyses of parameter sensitivity, convergence behavior, and representation visualization to provide more comprehensive insights into the effectiveness and stability of the proposed framework.

All revisions have been incorporated into the revised manuscript, with the corresponding changes highlighted in blue. Detailed explanations and responses to individual concerns are provided in our point-by-point rebuttal.

We sincerely hope that these revisions adequately address the reviewers' concerns and further clarify the methodological contributions and experimental effectiveness of TIGER. We would greatly appreciate any further feedback from the reviewers during the discussion period.

Thank you again for your time and consideration.

Sincerely,

The Authors of Submission

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.