acceptodds
Under review as a conference paper at ICLR 2027

Topology-Distilled Prompt Propagation for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to learn clusters from samples with partially observed views. Existing methods usually follow two directions, completing missing features or aligning the available views without imputation. Feature completion may inject biased synthetic signals because the missing view is not uniquely determined by the observed ones. Imputation-free methods avoid raw reconstruction, but they often rely on static consistency or pairwise alignment among observed views, which provides limited sample-specific guidance for refining missing-view representations. These limitations motivate a routing perspective in which the model learns how observed-view cues are encoded, exchanged, and fused across samples and missing patterns, rather than prescribing this process through a completion target or fixed alignment rule. We propose Topology-Distilled Prompt Propagation (TDPP), a framework that formulates missing-view representation learning as context-conditioned latent refinement. Its core module, cross-view prompt propagation, encodes observed-view cues into instance-specific prompts and iteratively exchanges them across views and samples under observation masks. This process enables missing-view representations to absorb reliable context while blocking messages from unobserved sources. To keep the propagated context consistent with the clustering objective, a self-supervised teacher trained on fully observed samples provides reference representations for the student. The teacher further guides the student with quality-aware feature distillation to account for unequal view reliability and sparse multi-hop topology distillation to preserve the neighborhood structure learned from complete data. Experiments on benchmark datasets show that TDPP outperforms recent state-of-the-art IMVC methods across missing ratios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Unsupervised Representation Learning, Incomplete Multi-View Clustering, Knowledge Distillation, Prompt Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rbp2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Topology-Distilled Prompt Propagation (TDPP) for incomplete multi-view clustering. A teacher trained on fully observed samples learns representations, prototypes, and neighborhood structure. The student replaces missing-view features with learnable placeholders and refines them through mask-aware cross-view prompt propagation (CVPP). Quality-aware feature distillation and sparse multi-hop topology distillation provide additional supervision. Experiments on four datasets show competitive results and substantial gains in some settings, notably on NUS-WIDE, although the method is not best on every metric.

Critical strengths

1.Refining missing-view representations in latent space, rather than reconstructing raw features, is a reasonable way to reduce dependence on potentially ambiguous imputation targets.

  1. Mask-aware prompt propagation and teacher-guided structural supervision form a coherent framework for handling incomplete observations.

  2. The reported clustering results are competitive, with sizeable improvements on NUS-WIDE at several missing ratios.

  3. The paper includes component ablations, parameter sensitivity analyses, and representation visualizations that offer useful evidence about the proposed design.

Critical weaknesses
  1. The novelty relative to directly related work is not adequately established. In particular, SP-IMVC (IEEE TIP 2026) also uses cross-view complementary prompts and global semantic guidance for incomplete multi-view clustering. The paper would benefit from a clearer mechanism-level comparison and a direct baseline. Related work on unequal view contributions is also relevant.

  2. All four datasets have only two views. Under the stated missingness protocol, each incomplete sample has exactly one observed view. Cross-sample attention is still active, but these experiments cannot test how CVPP combines several observed views or handles more complex missing patterns. The method also relies on complete samples for teacher training, while the effect of their scarcity remains untested despite the fallback in Eq. (18).

  3. Appendix B.3 reports a fixed random seed and averages metrics over the last three training epochs, rather than independent runs. This does not establish the stability of small gains under different initializations and missing masks. Some baselines were ported from MATLAB or adjusted to the common protocol, for which additional checks against the original implementations would strengthen the comparison.

  4. Several mechanisms need clearer justification. Appendix A states that the alignment loss in Eq. (A.5) directly updates the prototypes, not the teacher encoders; its contribution to learned representation quality is therefore unclear. The topology target in Eq. (15) includes the h = 0 identity term, but the influence of self-connections is not quantified. Cross-sample attention may also make representations sensitive to inference batch composition.

  5. Removing CVPP entirely shows that propagation matters, but does not establish an advantage over simpler, capacity-matched fusion or attention modules. Training cost, inference latency, and memory use are also not compared

Questions for authors
  1. How does TDPP differ from SP-IMVC in its prompting and structural guidance mechanisms? Could the authors include a direct comparison under the same evaluation protocol?

  2. How does TDPP perform with three or more views and heterogeneous missing patterns? As the complete subset shrinks, how do clustering performance and the validity indicator in Eq. (18) change?

  3. Since Eq. (A.5) directly updates only the prototypes, what evidence shows that this term improves the teacher’s clustering-relevant representations? Could the authors provide an ablation?

  4. In Eq. (15), how much of the teacher topology distribution is assigned to diagonal entries? What changes if the h = 0 identity term or self-connections are excluded?

  5. Are the student representations and final cluster assignments stable under different inference batch sizes and batch partitions?

  6. Could the authors report means and standard deviations over independent runs with varied seeds and missing masks, and explain the large dataset-dependent differences in the topology-distillation weight?

  7. How does CVPP compare with capacity-matched masked fusion or cross-view attention and with one-hop topology supervision, in both clustering quality and computational cost?

5 comments

We thank the reviewer for the careful reading of our paper and for the detailed and insightful comments! We respond to each point below.

Q1. How does TDPP differ from SP-IMVC in its prompting and structural guidance mechanisms? Could the authors include a direct comparison under the same evaluation protocol?

Response:

Thank you for this valuable suggestion. We have clarified the differences between TDPP and SP-IMVC in Section 2 of the revised manuscript. Specifically, the following discussion has been added and highlighted in blue:

SP-IMVC combines cross-view complementary prompts with latent anchor prompts to enrich observed-view representations and promote global semantic consistency.

TDPP focuses on refining missing-view representations through iterative mask-gated prompt propagation, guided by quality-aware feature distillation and sparse multi-hop topology distillation.

We have also included SP-IMVC as a direct baseline in Table 1 under our unified evaluation protocol. Since the datasets and missing-view mask generation protocols may differ from those used in the original SP-IMVC paper, some discrepancies from its originally reported results may occur.

Q2. How does TDPP perform with three or more views and heterogeneous missing patterns? As the complete subset shrinks, how do clustering performance and the validity indicator in Eq. (18) change?

Response:

Thank you for this constructive comment. We have extended our evaluation to OutdoorScene, which contains four distinct views, to examine the effectiveness of TDPP under more complex multi-view configurations. The results have been added to Table 9 and highlighted in blue, as shown below.

Table 9. Clustering performance on the four-view OutdoorScene dataset under different missing ratios. Results are reported as ACC/NMI/ARI (%).

Method0%20%50%80%
DIVIDE71.35/58.80/49.2770.70/57.67/48.4673.43/57.79/50.4663.68/50.45/40.70
CANDY78.06/63.76/57.7976.80/61.38/55.6372.61/58.01/50.6271.56/56.65/48.70
I2MVC68.82/54.11/45.9165.60/46.78/40.4859.23/39.35/32.6548.86/31.86/24.45
CAMEL54.86/39.82/31.4951.90/35.85/27.2742.26/30.08/19.7637.24/24.12/13.26
RAC-DMVC72.92/60.25/52.5673.80/60.36/52.3273.05/60.18/51.7171.65/59.18/51.40
GAVIM73.13/59.81/52.0065.14/56.88/46.9869.95/55.57/47.7164.68/52.94/43.49
DIMVC-HIA48.96/44.21/27.4865.61/49.13/38.0852.16/42.60/30.1752.84/39.11/26.09
SP-IMVC78.87/63.95/58.9372.92/56.41/50.2259.90/47.38/39.7644.90/35.23/27.42
TDPP91.17/82.66/80.9188.90/79.18/76.3786.35/77.09/71.9179.13/68.69/60.11

We have also evaluated TDPP under heterogeneous missingness patterns on NUS-WIDE, Scene15, and OutdoorScene at 20% and 50% missingness. Following the missing-pattern generation protocol of Missing Pattern Tree based Decision Grouping and Ensemble for Enhancing Pair Utilization in Deep Incomplete Multi-View Clustering (Xu et al., IEEE TPAMI, 2026, DOI: 10.1109/TPAMI.2026.3729698), we randomly select the specified proportion of samples as incomplete and mask between one and V−1V-1 entire views for each selected sample. This produces diverse sample-specific missing patterns, with at least one observed view retained per sample. The results indicate that TDPP maintains competitive clustering performance under these heterogeneous missingness patterns.

Table 10. Clustering performance under random and heterogeneous missingness patterns. Results are reported in percentage. Random missingness results are taken from the main table.

DatasetMissing rateMethodRandom missingness ACCRandom missingness NMIRandom missingness ARIHeterogeneous missingness ACCHeterogeneous missingness NMIHeterogeneous missingness ARI
NUS-WIDE20%DIMVC-HIA54.4141.4333.2054.3343.4834.96
NUS-WIDE20%TDPP70.8456.9753.3869.1456.7751.87
NUS-WIDE50%DIMVC-HIA51.4137.5728.7753.1039.1732.14
NUS-WIDE50%TDPP67.7854.9249.8870.0055.5051.95
Scene1520%DIMVC-HIA35.9537.9118.9036.7138.9919.57
Scene1520%TDPP54.4954.6337.5855.0455.0838.38
Scene1550%DIMVC-HIA36.8038.9219.8035.4936.6018.01
Scene1550%TDPP49.7750.3532.6348.2750.0032.19
OutdoorScene20%DIMVC-HIA65.6149.1338.0857.6846.5331.70
OutdoorScene20%TDPP88.9079.1876.3788.8079.5376.29
OutdoorScene50%DIMVC-HIA52.1642.6030.1753.3339.6527.19
OutdoorScene50%TDPP86.3577.0971.9186.6877.3372.28

The results indicate that TDPP maintains favorable clustering performance in the four-view setting and retains an advantage over the compared method under heterogeneous missingness. Some variations across missingness patterns are observed, suggesting that the configuration of available views can affect clustering performance.

As the complete subset shrinks, the teacher's clustering performance generally decreases, leading to a corresponding decline in student performance. The validity indicator η\eta in Eq. (18) is determined by the graph density of the frozen teacher on the complete subset. Topology supervision is enabled when the density is below 0.02 and disabled otherwise. A smaller complete subset may increase graph density and consequently deactivate topology supervision.

Q3. Since Eq. (A.5) directly updates only the prototypes, what evidence shows that this term improves the teacher's clustering-relevant representations? Could the authors provide an ablation?

Response:

Thank you for this constructive comment. We have conducted an additional ablation study comparing learned prototypes with directly computed centroids. The results have been incorporated into Section 4.3 and highlighted in blue, as shown below.

Table 11. Prototype Alignment Effectiveness Analysis at 50% missingness

DatasetLearned prototypes AA ACCLearned prototypes AA NMILearned prototypes AA ARIDirect centroids CC ACCDirect centroids CC NMIDirect centroids CC ARI
Scene1549.7750.3532.6349.4750.4032.49
NUS-WIDE67.7854.9249.8867.1154.5349.52
OutdoorScene86.3577.0971.9186.1476.8671.37

Learned prototypes provide modest improvements over directly computed centroids on most reported metrics. This suggests that prototype learning offers some benefit as a semantic reference for subsequent student prompt propagation.

We have also clarified the gradient flow and revised the discussion in Appendix A.5. Specifically, Eq. (A.5) updates only the learnable prototypes, without directly affecting the teacher encoder representations. The learned prototypes serve as class-level semantic references for subsequent student prompt propagation.

Q4. In Eq. (15), how much of the teacher topology distribution is assigned to diagonal entries? What changes if the h = 0 identity term or self-connections are excluded?

Response:

Thank you very much for this constructive and valuable comment. We quantify the average diagonal probability mass of the teacher topology distribution as

Dtch=1NC∑i=1NCp~tch(i∣i).D_{\mathrm{tch}} = \frac{1}{N_C}\sum_{i=1}^{N_C} \tilde{p}_{\mathrm{tch}}(i\mid i).

Under the original setting (H=2H=2), the diagonal entries receive 36.95% and 37.35% of the total probability mass on NUS-WIDE and Scene15, respectively.

We further investigated the effects of removing the h=0h=0 identity term and excluding self-connections. Removing h=0h=0 reduces the diagonal probability mass to 5.43% and 6.03%, respectively. Excluding self-connections entirely reduces it to zero. The corresponding clustering results are presented below.

Table 12. Effect of identity and self-connections on topology distillation at 50% missingness. DtchD_{\mathrm{tch}} and OtchO_{\mathrm{tch}} denote the mean diagonal and off-diagonal teacher probability masses. Results are reported in percentage.

DatasetVariantDtchD_{\mathrm{tch}}OtchO_{\mathrm{tch}}ACCNMIARI
NUS-WIDEFull36.9563.0567.7854.9249.88
NUS-WIDEw/o h=0h=05.4394.5768.0155.3350.36
NUS-WIDEw/o self-connections0.00100.0068.0455.4150.40
NUS-WIDEw/o topology----62.1848.8741.94
Scene15Full37.3562.6549.7750.3532.63
Scene15w/o h=0h=06.0393.9751.7951.6634.14
Scene15w/o self-connections0.00100.0051.1151.7833.99
Scene15w/o topology----44.3945.2327.20

Interestingly, removing either the identity term or self-connections improves clustering performance in these experiments. The results also indicate that topology distillation remains effective without self-matching, highlighting the importance of inter-sample relationships. We appreciate this valuable insight, which provides a promising direction for further refining our topology distillation strategy. We will investigate this aspect more thoroughly in future work.

Reviewer_PuN5OpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Topology-Distilled Prompt Propagation (TDPP) for incomplete multi-view clustering. A teacher trained on complete samples provides feature representations, view-quality estimates, and neighborhood topology. The student employs mask-gated cross-view prompt propagation and cross-sample attention to refine missing-view representations. Feature distillation, quality-aware distillation, and multi-hop topology distillation are jointly used to guide representation learning. Experiments on four datasets demonstrate competitive clustering performance under different missing ratios.

Critical strengths
  • The proposed mask-gated prompt propagation provides a reasonable mechanism for transferring contextual information from observed views to missing-view representations while preventing unavailable views from acting as information sources.

  • The teacher-student framework combines feature-level and topology-level supervision, providing complementary guidance for representation learning under incomplete observations.

  • The experimental results show competitive clustering performance. The ablation studies, parameter sensitivity analyses, and visualizations provide additional evidence for the effectiveness of the proposed framework.

Critical weaknesses
  • The topology distillation objective in Eq. (17) is defined only over complete samples, and Fig. 3 visualizes topology preservation on complete samples. These results do not directly establish whether the learned neighborhood structure is effectively transferred to genuinely incomplete samples.

  • The quality-aware weights in Eq. (13) regulate feature distillation but do not directly control the prompt routing in Eq. (4). Consequently, unreliable but observed views may still propagate misleading information, leaving the claimed reliability-aware information transfer insufficiently validated.

  • All four benchmark datasets contain only two views, and the evaluation considers randomly simulated missingness. The effectiveness under more complex view configurations and structured missing patterns remains unclear. The sensitivity to the number of complete samples used for teacher training is also unexplored.

  • The experiments use a single fixed random seed, without reporting results across multiple independent runs or standard deviations. Averaging performance over the final three epochs does not account for randomness in model initialization and missing-view masks, leaving the statistical reliability of the reported improvements unclear.

Questions for authors
  • Can the authors compare CVPP with capacity-matched attention fusion and evaluate quality-aware distillation against uniform weighting under matched supervision strength?

  • Can the authors provide quantitative evidence that topology distillation improves neighborhood preservation specifically for incomplete samples?

  • How does the framework prevent unreliable observed views from propagating misleading information when quality weights do not directly affect prompt routing?

  • Can the authors evaluate more complex view configurations, structured missing patterns, and different proportions of complete samples for teacher training?

  • Can the authors report clustering results over multiple independent runs with different random seeds, including mean values and standard deviations?

3 comments

We thank the reviewer for the insightful feedback and valuable comments! Below, we address your comments and suggestions in detail.

Q1. Can the authors compare CVPP with capacity-matched attention fusion and evaluate quality-aware distillation against uniform weighting under matched supervision strength?

Response:

Thank you for this valuable suggestion. To provide a more rigorous comparison, we have introduced a parameter-size-matched cross-view attention baseline (PS-CVA), which replaces CVPP with conventional masked cross-view attention while retaining the same student backbone, teacher supervision, and training objectives. The interaction parameters are closely matched to ensure a fair comparison. The effectiveness of quality-aware distillation is also evaluated in the ablation study in Section 4.3. The comparison results are presented below.

Table 3. Comparison with parameter-size-matched cross-view attention at 50% missingness. Interaction parameters exclude the shared student backbone and teacher. Clustering results are reported in percentage.

DatasetMethodInteraction parametersACCNMIARI
NUS-WIDETDPP363,77667.7854.9249.88
NUS-WIDEPS-CVA362,94463.1448.3142.58
OutdoorSceneTDPP396,67286.3577.0971.91
OutdoorScenePS-CVA395,90477.1661.4256.02

Under closely matched interaction parameter counts, TDPP achieves better clustering performance on both datasets. This suggests that the improvements are not solely attributable to model capacity, supporting the benefit of the proposed prompt propagation mechanism.

Q2. Can the authors provide quantitative evidence that topology distillation improves neighborhood preservation specifically for incomplete samples?

Response:

We thank the reviewer for this insightful comment. To quantitatively evaluate its effectiveness on incomplete samples, we introduce Neighborhood Purity@10, which measures the local semantic consistency of incomplete representations. For each incomplete sample ii, we retrieve its ten nearest neighbors N10(i)\mathcal{N}_{10}(i) from all samples using cosine similarity between the normalized concatenated student representations, excluding the query itself. The metric is defined as

P@10 = \frac{1}{|\mathcal{D}_I|} \sum_{i \in \mathcal{D}_I} \frac{1}{10} \sum_{j \in \mathcal{N}_{10}(i)} \mathbb{I}[y_i = y_j], \tag{1}

where DI\mathcal{D}_I denotes the set of incomplete samples and yiy_i is the ground-truth class label. Labels are used only for evaluation, not for neighbor retrieval or model training.

We compare the full TDPP model with its variant without topology distillation under 50% missingness. The quantitative results are presented below.

Table 4. Neighborhood Purity@10 of incomplete samples at 50% missingness. Neighbors are retrieved from all samples using cosine similarity. Δ\Delta denotes the improvement in percentage points.

Dataset# Incompletew/o Topology (%)Full (%)Δ\Delta (pp)
Scene152,24354.2258.55+4.33
NUS-WIDE4,50059.7566.33+6.58

As shown above, topology distillation improves Neighborhood Purity@10 by 4.33 and 6.58 percentage points on Scene15 and NUS-WIDE, respectively. These improvements indicate that incomplete samples acquire more semantically consistent local neighborhoods through topology-guided student learning. Although structural supervision is constructed from complete samples, its benefits extend to incomplete representations, supporting the effectiveness of topology distillation in preserving clustering-relevant neighborhood relationships.

Q3. How does the framework prevent unreliable observed views from propagating misleading information when quality weights do not directly affect prompt routing?

Response:

We thank the reviewer for this valuable comment. We would like to clarify that quality-aware distillation and mask-gated prompt routing serve complementary purposes. The quality weights in Eq. (13) are estimated from complete samples and adaptively regulate feature-level supervision according to each view's local neighborhood reliability. This encourages the student to learn from more reliable representations without directly modifying the routing weights. In contrast, the observation masks in Eqs. (4) and (9) ensure that only observed views can propagate contextual prompts, whereas missing views can receive information but cannot act as sources. We acknowledge that this mechanism does not explicitly suppress unreliable observed views during inference. Instead, their influence is addressed indirectly through quality-aware supervision during training.

Q4. Can the authors evaluate more complex view configurations, structured missing patterns, and different proportions of complete samples for teacher training?

Response (Part 1/2):

Thank you for this constructive suggestion. We have extended our evaluation to OutdoorScene, which contains four distinct views, to examine the effectiveness of TDPP under more complex multi-view configurations. The additional results have been incorporated into Table 5 and highlighted in blue.

Table 5. Clustering performance on the four-view OutdoorScene dataset under different missing ratios. Results are reported as ACC/NMI/ARI (%).

Method0%20%50%80%
DIVIDE71.35/58.80/49.2770.70/57.67/48.4673.43/57.79/50.4663.68/50.45/40.70
CANDY78.06/63.76/57.7976.80/61.38/55.6372.61/58.01/50.6271.56/56.65/48.70
I2MVC68.82/54.11/45.9165.60/46.78/40.4859.23/39.35/32.6548.86/31.86/24.45
CAMEL54.86/39.82/31.4951.90/35.85/27.2742.26/30.08/19.7637.24/24.12/13.26
RAC-DMVC72.92/60.25/52.5673.80/60.36/52.3273.05/60.18/51.7171.65/59.18/51.40
GAVIM73.13/59.81/52.0065.14/56.88/46.9869.95/55.57/47.7164.68/52.94/43.49
DIMVC-HIA48.96/44.21/27.4865.61/49.13/38.0852.16/42.60/30.1752.84/39.11/26.09
SP-IMVC78.87/63.95/58.9372.92/56.41/50.2259.90/47.38/39.7644.90/35.23/27.42
TDPP91.17/82.66/80.9188.90/79.18/76.3786.35/77.09/71.9179.13/68.69/60.11

The results indicate that TDPP maintains clustering performance across different missing ratios in the four-view setting.

We further evaluate TDPP under heterogeneous missingness patterns on NUS-WIDE, Scene15, and OutdoorScene at 20% and 50% missingness. Following the missing-pattern generation protocol of Missing Pattern Tree based Decision Grouping and Ensemble for Enhancing Pair Utilization in Deep Incomplete Multi-View Clustering (Xu et al., IEEE TPAMI, 2026, DOI: 10.1109/TPAMI.2026.3729698), we randomly select the specified proportion of samples as incomplete and mask between one and V−1V-1 entire views for each selected sample. This produces diverse sample-specific missing patterns, with at least one observed view retained per sample. The results indicate that TDPP maintains competitive clustering performance under these heterogeneous missingness patterns.

Table 6. Clustering performance under random and heterogeneous missingness patterns. Results are reported in percentage. Random missingness results are taken from the main table.

DatasetMissing rateMethodRandom missingness ACCRandom missingness NMIRandom missingness ARIHeterogeneous missingness ACCHeterogeneous missingness NMIHeterogeneous missingness ARI
NUS-WIDE20%DIMVC-HIA54.4141.4333.2054.3343.4834.96
NUS-WIDE20%TDPP70.8456.9753.3869.1456.7751.87
NUS-WIDE50%DIMVC-HIA51.4137.5728.7753.1039.1732.14
NUS-WIDE50%TDPP67.7854.9249.8870.0055.5051.95
Scene1520%DIMVC-HIA35.9537.9118.9036.7138.9919.57
Scene1520%TDPP54.4954.6337.5855.0455.0838.38
Scene1550%DIMVC-HIA36.8038.9219.8035.4936.6018.01
Scene1550%TDPP49.7750.3532.6348.2750.0032.19
OutdoorScene20%DIMVC-HIA65.6149.1338.0857.6846.5331.70
OutdoorScene20%TDPP88.9079.1876.3788.8079.5376.29
OutdoorScene50%DIMVC-HIA52.1642.6030.1753.3339.6527.19
OutdoorScene50%TDPP86.3577.0971.9186.6877.3372.28

TDPP maintains better clustering performance than the compared method under heterogeneous missingness across the evaluated datasets. The results vary moderately across missingness patterns, suggesting that the proposed framework can accommodate different combinations of observed views.

Furthermore, we have examined the dependence on the complete subset by reporting teacher and student clustering performance under different missing ratios. The proportions of complete samples available for teacher pretraining are 100%, 80%, 50%, and 20%, respectively. The student architecture and training budget remain unchanged. As shown below, the teacher's clustering performance generally decreases as the complete subset shrinks. The student also exhibits some performance degradation but maintains competitive clustering performance under severe incompleteness.

Response (Q4, Part 2/2):

Table 7. Teacher and student clustering performance on NUS-WIDE and OutdoorScene under different missing ratios. Results are reported in percentage.

DatasetMissing ratioTeacher ACCTeacher NMITeacher ARIStudent ACCStudent NMIStudent ARI
NUS-WIDE0%59.666750.984043.565372.644460.259856.0446
NUS-WIDE20%63.677850.336842.351870.766756.847053.2637
NUS-WIDE50%60.544448.156842.381669.488954.529350.8960
NUS-WIDE80%43.066735.786222.975864.677849.516944.9449
OutdoorScene0%79.427164.312259.688991.294682.770781.1467
OutdoorScene20%75.892960.827754.281688.950980.181976.5580
OutdoorScene50%75.706859.669153.794186.904877.258572.7762
OutdoorScene80%65.513450.041540.851879.389970.812560.7191

Q5. Can the authors report clustering results over multiple independent runs with different random seeds, including mean values and standard deviations?

Response:

Thank you for this valuable suggestion. We have conducted additional experiments with multiple independent random seeds across five datasets and four missing ratios. The results, including mean values and sample variances, have been added to Appendix C.1 and highlighted in blue, as shown below.

Table 8. Supplementary multi-seed clustering results. Entries report mean ±\pm sample variance (ddof=1\mathrm{ddof}=1), with means in percent and variances in squared percentage points.

DatasetMetricMissing rate 0%Missing rate 20%Missing rate 50%Missing rate 80%
CUBACC83.24±0.09583.24 \pm 0.09582.72±0.10782.72 \pm 0.10775.72±0.40975.72 \pm 0.40970.54±0.03670.54 \pm 0.036
CUBNMI78.81±0.13578.81 \pm 0.13576.70±0.93476.70 \pm 0.93470.03±0.30570.03 \pm 0.30565.06±0.37965.06 \pm 0.379
CUBARI69.97±0.30269.97 \pm 0.30268.49±0.92168.49 \pm 0.92158.87±0.07558.87 \pm 0.07552.25±0.29052.25 \pm 0.290
NUS-WIDEACC72.48±0.58972.48 \pm 0.58969.74±1.00869.74 \pm 1.00868.66±0.70968.66 \pm 0.70964.06±4.28964.06 \pm 4.289
NUS-WIDENMI59.83±1.09459.83 \pm 1.09456.63±0.08756.63 \pm 0.08754.21±0.47154.21 \pm 0.47149.14±1.13449.14 \pm 1.134
NUS-WIDEARI55.75±2.09055.75 \pm 2.09051.88±2.02151.88 \pm 2.02149.90±1.02749.90 \pm 1.02744.28±3.19844.28 \pm 3.198
Scene15ACC56.38±0.11456.38 \pm 0.11453.51±0.80953.51 \pm 0.80949.76±0.01149.76 \pm 0.01144.56±1.93144.56 \pm 1.931
Scene15NMI57.08±0.10357.08 \pm 0.10354.12±0.19554.12 \pm 0.19550.06±0.08250.06 \pm 0.08244.67±3.06844.67 \pm 3.068
Scene15ARI39.83±1.27139.83 \pm 1.27137.17±0.26037.17 \pm 0.26032.55±0.00532.55 \pm 0.00526.59±3.92226.59 \pm 3.922
LandUse-21ACC31.10±0.02031.10 \pm 0.02030.51±0.17630.51 \pm 0.17630.06±0.03530.06 \pm 0.03529.43±0.16029.43 \pm 0.160
LandUse-21NMI36.94±0.01236.94 \pm 0.01235.83±0.04735.83 \pm 0.04734.43±0.97734.43 \pm 0.97734.00±0.47734.00 \pm 0.477
LandUse-21ARI16.48±0.02716.48 \pm 0.02715.96±0.02015.96 \pm 0.02015.47±0.07115.47 \pm 0.07115.32±0.05315.32 \pm 0.053
OutdoorSceneACC90.72±0.33890.72 \pm 0.33889.05±0.05889.05 \pm 0.05886.08±2.03586.08 \pm 2.03580.00±0.67780.00 \pm 0.677
OutdoorSceneNMI82.01±0.83182.01 \pm 0.83179.77±0.26679.77 \pm 0.26676.93±2.48576.93 \pm 2.48569.48±1.04769.48 \pm 1.047
OutdoorSceneARI80.01±1.32480.01 \pm 1.32476.69±0.19876.69 \pm 0.19871.39±5.98771.39 \pm 5.98761.20±1.21661.20 \pm 1.216

The results show broadly consistent performance across multiple runs, with some variation depending on the dataset and missingness level.

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

This paper studies incomplete multi-view clustering and proposes TDPP, a teacher–student framework. The teacher learns representations, prototypes, and neighborhood structure from complete samples. The student updates missing-view representations through prompt propagation with observation masks and receives feature and topology supervision from the teacher. The authors aim to avoid directly reconstructing missing raw features while preserving structural information useful for clustering. The method achieves good results on some datasets, but the experiments have not established its advantage over conventional feature interaction or adequately examined whether it remains effective when few complete samples are available.

Critical strengths
  1. Allowing missing views to receive context while preventing them from serving as information sources is a reasonable choice. The role of the observation masks in propagation is also clear.

  2. Quality-aware distillation adjusts supervision weights according to how well each view preserves local neighborhoods, accounting for the fact that a view’s reliability may vary across samples. This design can help reduce interference from weaker views during distillation and is reasonably motivated.

  3. The improvements on some datasets are substantial, and the ablation results indicate that the teacher’s topology information benefits the current model. These findings make the overall approach worth further investigation.

Critical weaknesses
  1. The teacher, prototypes, and topology supervision all depend on the complete subset, but the paper does not examine cases where complete samples are scarce or do not adequately cover the classes. The topology visualizations also include only complete samples, leaving insufficient evidence that the teacher’s structure transfers effectively to incomplete samples.

  2. Removing the entire CVPP module also removes the main channel through which missing-view representations obtain information. The resulting performance drop therefore does not establish the value of the prompt design itself. The paper lacks replacement baselines using standard masked attention or direct feature transmission, so the independent contribution of prompt propagation remains unclear.

  3. The paper does not specify how samples are batched during inference or when topology supervision is enabled, which affects reproducibility. The experiments also lack repeated runs with different random seeds, and some metrics do not outperform competing methods. The current claims about stability and performance advantages therefore appear too strong.

Questions for authors
  1. Does the method remain effective when the teacher has fewer complete samples or when those samples fail to cover some classes? I suggest varying the number of complete samples available to the teacher while keeping the student architecture and training budget unchanged, and reporting performance separately for complete and incomplete samples to examine the method’s dependence on the complete subset.

  2. How are attention reference samples selected during inference, and are the results stable under different batching schemes? Please explain how cases with no available keys or values are handled. For reproducibility, please also clarify whether topology weights are selected without ground-truth labels, the activation rule in Eq. (18), and the actual settings used in each experiment.

  3. Appendix A.5 updates only the prototypes and stops gradients through the centroids. How, then, does this objective affect the clustering structure of the teacher encoder? Please clarify the gradient flow. If this term merely fits prototypes to centroids, I suggest revising the discussion of semantic alignment to avoid equating prototype updates with improvements in the teacher representations.

3 comments

Thank you for your insightful comments and for taking the time to review our paper. We respond to your suggestions point by point as follows.

Q1. Does the method remain effective when the teacher has fewer complete samples or when those samples fail to cover some classes? I suggest varying the number of complete samples available to the teacher while keeping the student architecture and training budget unchanged, and reporting performance separately for complete and incomplete samples to examine the method's dependence on the complete subset.

Response:

Thank you for this valuable comment. We would like to clarify that the teacher is pretrained exclusively on fully observed samples under each corresponding missingness setting, rather than on an independently available complete dataset. Consequently, the proportions of complete samples available for teacher training are 100%, 80%, 50%, and 20% under missing rates of 0%, 20%, 50%, and 80%, respectively.

To further examine this dependence, we have supplemented the teacher and student clustering results on NUS-WIDE and OutdoorScene across different missing rates, keeping the student architecture and training budget unchanged. As shown below, the teacher's clustering performance generally declines as the complete subset shrinks. Nevertheless, the student maintains competitive performance even under severe missingness, achieving 64.68% and 79.39% ACC on NUS-WIDE and OutdoorScene at 80% missingness, respectively. The additional results and corresponding discussion have been highlighted in blue.

Table 1. Teacher and student clustering performance on NUS-WIDE and OutdoorScene under different missing ratios. Results are reported in percentage.

DatasetMissing ratioTeacher ACCTeacher NMITeacher ARIStudent ACCStudent NMIStudent ARI
NUS-WIDE0%59.666750.984043.565372.644460.259856.0446
NUS-WIDE20%63.677850.336842.351870.766756.847053.2637
NUS-WIDE50%60.544448.156842.381669.488954.529350.8960
NUS-WIDE80%43.066735.786222.975864.677849.516944.9449
OutdoorScene0%79.427164.312259.688991.294682.770781.1467
OutdoorScene20%75.892960.827754.281688.950980.181976.5580
OutdoorScene50%75.706859.669153.794186.904877.258572.7762
OutdoorScene80%65.513450.041540.851879.389970.812560.7191

TDPP does not strictly require the complete subset to cover all classes. Although the teacher provides structural guidance from complete samples, the student processes all samples and refines their representations through mask-gated prompt propagation and cross-sample attention over observed views. Moreover, final clustering is performed on all student representations without restricting cluster assignments to the teacher prototypes. Thus, samples from classes absent in the complete subset can still be represented and clustered, although their performance may be affected by the limited teacher guidance.

Q2. How are attention reference samples selected during inference, and are the results stable under different batching schemes? Please explain how cases with no available keys or values are handled. For reproducibility, please also clarify whether topology weights are selected without ground-truth labels, the activation rule in Eq. (18), and the actual settings used in each experiment.

Response:

Thank you for these valuable questions. We provide the following clarifications.

Attention reference selection and Inference batch sensitivity During inference, the trained student and CVPP parameters are fixed. Within each inference batch, samples with observed target views serve as attention keys and values, and all samples can act as queries. Our missingness protocol ensures that each sample retains at least one observed view. If no valid keys or values are available for a particular view, the corresponding attention operation is simply skipped.

We have evaluated different inference batch sizes and sample partitions on three datasets at 20% missingness, keeping the training batch size fixed at 256. The results exhibit sensitivity to inference batch size and partitioning, but TDPP consistently maintains competitive clustering performance across the evaluated settings. The additional experiments have been incorporated into Section 4.2 and highlighted in blue. The results are presented below.

Table 2. Clustering performance under different inference batch sizes and partitions at 20% missingness. The training batch size is 256. Results are reported as the mean over three random seeds, with means in percentage.

Mean (%) ±\pm variance

Inference Batch SizeScene15 ACCScene15 NMIScene15 ARINUS-WIDE ACCNUS-WIDE NMINUS-WIDE ARIOutdoorScene ACCOutdoorScene NMIOutdoorScene ARI
6455.11±0.05155.11 \pm 0.05155.56±0.02155.56 \pm 0.02137.90±0.07337.90 \pm 0.07369.17±0.00969.17 \pm 0.00955.90±0.01155.90 \pm 0.01151.94±0.01851.94 \pm 0.01894.13±0.00994.13 \pm 0.00987.84±0.05787.84 \pm 0.05786.81±0.04186.81 \pm 0.041
12853.62±3.45653.62 \pm 3.45654.70±0.20754.70 \pm 0.20736.68±1.33836.68 \pm 1.33869.14±0.11869.14 \pm 0.11855.79±0.09255.79 \pm 0.09251.88±0.25351.88 \pm 0.25392.77±0.10692.77 \pm 0.10685.59±0.42185.59 \pm 0.42184.08±0.44984.08 \pm 0.449
25653.79±0.01553.79 \pm 0.01554.35±0.03754.35 \pm 0.03736.45±0.03936.45 \pm 0.03968.96±0.17468.96 \pm 0.17455.37±0.00255.37 \pm 0.00251.41±0.02951.41 \pm 0.02989.61±0.23689.61 \pm 0.23680.40±0.49580.40 \pm 0.49577.73±0.81577.73 \pm 0.815
51252.58±0.20052.58 \pm 0.20053.14±0.04953.14 \pm 0.04934.98±0.11734.98 \pm 0.11768.03±0.00668.03 \pm 0.00654.58±0.01754.58 \pm 0.01750.31±0.02450.31 \pm 0.02486.27±0.02686.27 \pm 0.02675.28±0.01375.28 \pm 0.01371.71±0.07671.71 \pm 0.076

Topology weight selection and Activation rule and experimental settings The topology supervision is constructed from the teacher representations without ground-truth labels. The topology weights are configured according to dataset characteristics. For smaller datasets, the limited complete subset may yield less reliable neighborhood relationships, and excessive topology supervision could blur inter-cluster boundaries.

The validity indicator η\eta in Eq. (18) is determined from the graph constructed using the frozen teacher representations on the complete subset. Topology supervision is activated (η=1\eta=1) when the graph density is below 0.02 and disabled (η=0\eta=0) otherwise. The dataset-specific hyperparameters and training configurations are reported in Appendix B.3.

Q3. Appendix A.5 updates only the prototypes and stops gradients through the centroids. How, then, does this objective affect the clustering structure of the teacher encoder? Please clarify the gradient flow. If this term merely fits prototypes to centroids, I suggest revising the discussion of semantic alignment to avoid equating prototype updates with improvements in the teacher representations.

Response:

Thank you for this insightful comment. We agree that the alignment objective in Eq. (A.5) updates only the learnable prototypes, with gradients stopped through the centroid targets, and does not directly improve the teacher encoder representations. We have clarified the gradient flow and refined the discussion of semantic alignment in Appendix A.5. The revised text, highlighted in blue, reads as follows:

"Because all blocks Cvv=1V\\{C^v\\}_{v=1}^{V} originate from the same centroid in the concatenated space, the learned prototypes preserve consistent cluster indices across views. This objective updates only the prototypes, without directly affecting the teacher representations. The resulting prototypes serve as class-level semantic references for subsequent student prompt propagation."

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Updates

Dear PCs, SACs, ACs, and Reviewers,

We sincerely thank all reviewers for their constructive feedback and valuable suggestions, which have helped us improve our manuscript.

In response to the reviewers' concerns, we have made substantial efforts to conduct additional experiments, perform further analyses, and revise the manuscript. The major updates are summarized below:

  • We added the SP-IMVC baseline, parameter-size-matched cross-view attention comparisons, and further topology distillation ablations.
  • We extended evaluations to the four-view OutdoorScene dataset, heterogeneous missingness patterns, and varying proportions of complete samples.
  • We examined inference batch sensitivity and conducted multi-seed experiments across five datasets and four missing ratios.
  • We added neighborhood preservation analyses, prototype alignment ablations, and clarifications on gradient flow, attention routing, and topology supervision.

All updates have been incorporated into the revised manuscript, with changes highlighted in blue. Detailed explanations are provided in our point-by-point responses.

We hope these efforts help address the reviewers' concerns and clarify our contributions. 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 #8230

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.