- 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.
-
Mask-aware prompt propagation and teacher-guided structural supervision form a coherent framework for handling incomplete observations.
-
The reported clustering results are competitive, with sizeable improvements on NUS-WIDE at several missing ratios.
-
The paper includes component ablations, parameter sensitivity analyses, and representation visualizations that offer useful evidence about the proposed design.
Critical weaknesses
-
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.
-
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).
-
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.
-
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.
-
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
-
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?
-
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?
-
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?
-
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?
-
Are the student representations and final cluster assignments stable under different inference batch sizes and batch partitions?
-
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?
-
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
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 (%).
| Method | 0% | 20% | 50% | 80% |
|---|---|---|---|---|
| DIVIDE | 71.35/58.80/49.27 | 70.70/57.67/48.46 | 73.43/57.79/50.46 | 63.68/50.45/40.70 |
| CANDY | 78.06/63.76/57.79 | 76.80/61.38/55.63 | 72.61/58.01/50.62 | 71.56/56.65/48.70 |
| I2MVC | 68.82/54.11/45.91 | 65.60/46.78/40.48 | 59.23/39.35/32.65 | 48.86/31.86/24.45 |
| CAMEL | 54.86/39.82/31.49 | 51.90/35.85/27.27 | 42.26/30.08/19.76 | 37.24/24.12/13.26 |
| RAC-DMVC | 72.92/60.25/52.56 | 73.80/60.36/52.32 | 73.05/60.18/51.71 | 71.65/59.18/51.40 |
| GAVIM | 73.13/59.81/52.00 | 65.14/56.88/46.98 | 69.95/55.57/47.71 | 64.68/52.94/43.49 |
| DIMVC-HIA | 48.96/44.21/27.48 | 65.61/49.13/38.08 | 52.16/42.60/30.17 | 52.84/39.11/26.09 |
| SP-IMVC | 78.87/63.95/58.93 | 72.92/56.41/50.22 | 59.90/47.38/39.76 | 44.90/35.23/27.42 |
| TDPP | 91.17/82.66/80.91 | 88.90/79.18/76.37 | 86.35/77.09/71.91 | 79.13/68.69/60.11 |
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 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.
| Dataset | Missing rate | Method | Random missingness ACC | Random missingness NMI | Random missingness ARI | Heterogeneous missingness ACC | Heterogeneous missingness NMI | Heterogeneous missingness ARI |
|---|---|---|---|---|---|---|---|---|
| NUS-WIDE | 20% | DIMVC-HIA | 54.41 | 41.43 | 33.20 | 54.33 | 43.48 | 34.96 |
| NUS-WIDE | 20% | TDPP | 70.84 | 56.97 | 53.38 | 69.14 | 56.77 | 51.87 |
| NUS-WIDE | 50% | DIMVC-HIA | 51.41 | 37.57 | 28.77 | 53.10 | 39.17 | 32.14 |
| NUS-WIDE | 50% | TDPP | 67.78 | 54.92 | 49.88 | 70.00 | 55.50 | 51.95 |
| Scene15 | 20% | DIMVC-HIA | 35.95 | 37.91 | 18.90 | 36.71 | 38.99 | 19.57 |
| Scene15 | 20% | TDPP | 54.49 | 54.63 | 37.58 | 55.04 | 55.08 | 38.38 |
| Scene15 | 50% | DIMVC-HIA | 36.80 | 38.92 | 19.80 | 35.49 | 36.60 | 18.01 |
| Scene15 | 50% | TDPP | 49.77 | 50.35 | 32.63 | 48.27 | 50.00 | 32.19 |
| OutdoorScene | 20% | DIMVC-HIA | 65.61 | 49.13 | 38.08 | 57.68 | 46.53 | 31.70 |
| OutdoorScene | 20% | TDPP | 88.90 | 79.18 | 76.37 | 88.80 | 79.53 | 76.29 |
| OutdoorScene | 50% | DIMVC-HIA | 52.16 | 42.60 | 30.17 | 53.33 | 39.65 | 27.19 |
| OutdoorScene | 50% | TDPP | 86.35 | 77.09 | 71.91 | 86.68 | 77.33 | 72.28 |
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 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
| Dataset | Learned prototypes ACC | Learned prototypes NMI | Learned prototypes ARI | Direct centroids ACC | Direct centroids NMI | Direct centroids ARI |
|---|---|---|---|---|---|---|
| Scene15 | 49.77 | 50.35 | 32.63 | 49.47 | 50.40 | 32.49 |
| NUS-WIDE | 67.78 | 54.92 | 49.88 | 67.11 | 54.53 | 49.52 |
| OutdoorScene | 86.35 | 77.09 | 71.91 | 86.14 | 76.86 | 71.37 |
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
Under the original setting (), 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 identity term and excluding self-connections. Removing 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. and denote the mean diagonal and off-diagonal teacher probability masses. Results are reported in percentage.
| Dataset | Variant | ACC | NMI | ARI | ||
|---|---|---|---|---|---|---|
| NUS-WIDE | Full | 36.95 | 63.05 | 67.78 | 54.92 | 49.88 |
| NUS-WIDE | w/o | 5.43 | 94.57 | 68.01 | 55.33 | 50.36 |
| NUS-WIDE | w/o self-connections | 0.00 | 100.00 | 68.04 | 55.41 | 50.40 |
| NUS-WIDE | w/o topology | -- | -- | 62.18 | 48.87 | 41.94 |
| Scene15 | Full | 37.35 | 62.65 | 49.77 | 50.35 | 32.63 |
| Scene15 | w/o | 6.03 | 93.97 | 51.79 | 51.66 | 34.14 |
| Scene15 | w/o self-connections | 0.00 | 100.00 | 51.11 | 51.78 | 33.99 |
| Scene15 | w/o topology | -- | -- | 44.39 | 45.23 | 27.20 |
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.
We thank the reviewer for the careful reading of our paper and for the detailed and insightful comments! We respond to each point below.
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.