- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies incomplete multi-view clustering, focusing on how to recover a shared cluster structure when some views are missing. Existing methods face three main issues: borrowed graph topologies are unreliable, high-confidence predictions are not well calibrated, and guidance signals conflict with contrastive learning. To address these problems, the paper introduces the RTG-IMVC framework. It builds candidate topology sets and computes sample-level transfer risk certificates to handle unreliable topology transfer. It then uses a logarithmic opinion pool with inverse-risk weights to combine expert predictions, which keeps high-risk experts from misleading the model. Finally, it uses a minimum-change projection to ensure the base contrastive learning objective does not get worse in a single step, removing the conflict between guidance signals and representation learning.
Critical strengths
The combination of risk-normalized LogOP and base-safe projection is novel. Converting prediction variation under topology perturbations into expert weights, and using a half-space projection to protect the main contrastive objective, forms a coherent chain: structural uncertainty → consensus formation → optimization safety.
Critical weaknesses
- The risk is the prediction variation of the same encoder across candidate topologies, which is not equivalent to the reliability of the transferred topology. Moreover, no estimation procedure or sensitivity analysis is given for (L) and (epsilon).
- The fusion operator, (mathcal{L}{ins}), and (mathcal{L}{clu}) are undefined; there is no Big-O complexity analysis; each missing sample requires K candidate-topology forward passes and dual gradients, but GPU memory and large-graph scalability are not discussed. In addition, the base-safe projection may protect an incorrect base objective; this optimization conservatism is not analyzed.
Questions for authors
-
How are and determined in experiments? Are they global constants, per-sample dynamic estimates, or tuned on validation data? Please provide specific values or the estimation algorithm, and add a sensitivity analysis for underestimation/overestimation of .
-
Does measure the reliability of the transferred topology itself or the sensitivity of the encoder to topology perturbations? If the encoder is very robust, can a wrong topology still receive low risk? Please clarify conceptually and provide a diagnostic experiment that distinguishes the two.
-
What is the fusion operator for (attention, concatenation, averaging, etc.)? Please write out explicit formulas for and .
-
How often are the detached risks, weights, and teacher refreshed? Could detachment cause a mismatch between the risk estimate and the evolving model? Is there theoretical justification for detaching rather than differentiating through the risk?
-
How often does the projection actually activate (fraction of )? What is the distribution of cosine similarity between and ? Is the gain from Step 3 to Step 4 statistically significant across ten seeds?
-
Please provide per-iteration time complexity and GPU memory curves as functions of sample size, number of views, and candidate number . Does the framework support mini-batch subgraph sampling, or does it require full-batch graph processing?
-
How does functionally differ from the conflict-avoidance mechanisms of PCGrad, CAGrad, and MGDA? Is it a relaxation of these methods or a fundamentally different constraint?
-
At , ACC on NoisyMNIST decreases from at epoch 100 to at epoch 500. Does this indicate that the base-safe projection is protecting an incorrect base objective? Would the authors consider periodically "releasing" the constraint or introducing a correction mechanism?
4 comments
Q2. Does measure the reliability of the transferred topology itself, or the encoder's sensitivity to it?
By design, measures the latter: it bounds how much an expert's log-prediction can change over plausible topologies (Theorem 1). In other words, it certifies the decision rather than the graph. This is the quantity the teacher needs, because Theorem 2 bounds how far a perturbed expert can move the teacher's log-odds. In the reviewer's scenario (a very robust encoder with a wrong topology), is indeed low, but so is the harm: if the expert's prediction does not change across plausible topologies, a wrong topology cannot change the teacher either. We added this clarification to Sec. 3.2.
To separate the two notions, we corrupt 30% of the transferred rows by replacing half of their neighbors with other-class samples (labels are used only to build this diagnostic) and vary the propagation weight that controls how strongly the encoder uses topology (: topology-free; : RTG-IMVC). The protocol is in App. H.1.
Table 3. Risk diagnostic on NoisyMNIST, . ΔACC is the accuracy change caused by the corruption; AUROC measures how well a score detects corrupted rows (PDF: Table 6(a)).
| ACC (clean) | ΔACC (ours) | ΔACC (no certificate) | AUROC of | AUROC of confidence | Spearman(, neighbor error) | |
|---|---|---|---|---|---|---|
| 0 (topology-free) | 86.41 | −0.14 | −0.21 | 0.51 | 0.53 | 0.03 |
| 0.5 | 91.86 | −1.74 | −4.95 | 0.79 | 0.57 | 0.48 |
| 1 (default) | 93.30 | −1.83 | −6.62 | 0.87 | 0.59 | 0.62 |
Table 4. Risk quintiles on clean NoisyMNIST, (PDF: Table 6(b) and Table R4).
| Q1 (low ) | Q2 | Q3 | Q4 | Q5 (high ) | |
|---|---|---|---|---|---|
| Neighbor purity (%) | 94.8 | 90.3 | 83.6 | 72.9 | 55.4 |
| Teacher accuracy (%) | 98.1 | 96.4 | 93.2 | 87.5 | 76.0 |
| Weight of the transferred expert | 0.29 | 0.23 | 0.18 | 0.13 | 0.09 |
| Guidance weight | 0.88 | 0.73 | 0.59 | 0.47 | 0.37 |
Table 3 shows the two regimes. With , detects corrupted rows far better than prediction confidence, and the accuracy loss is limited. With , , and the corruption leaves ACC essentially unchanged, which is the intended behavior. Table 4 shows that higher-risk pairs have less pure transferred neighborhoods and a less accurate teacher, and receive smaller weights. We also state the remaining limitation explicitly (App. F.1, Limitation 3): if all donor views share the same wrong neighbors, topology variation alone cannot reveal it.
Q3. Fusion operator for and explicit , .
The fused prediction uses instance-level attention fusion over all view embeddings of sample (observed or transferred) (App. G.2, Eq. 45):
With batch , cosine similarity , cluster-assignment columns , and (App. G.3, Eq. 46):
Missing pairs enter both losses through their nominal transferred embedding.
Q4. Refresh frequency of the detached quantities; justification for detaching.
- Frequency. Risks, weights, and teacher are refreshed at every iteration (Sec. 3.5, Algorithm 2).
- Why detach. Since and decreases in every risk, both and are non-increasing in every risk, and so is the guidance term. If gradients flowed through the risks, descent would push the network to increase its own risk (i.e., make predictions more topology-sensitive) so that the guidance switches itself off. Detaching removes this shortcut, as in standard fixed-target self-training (App. G.5).
- Mismatch. Remark 2 (App. G.5) shows that between consecutive refreshes the diameter changes by at most , so with per-iteration refresh the staleness is first order in the step size.
Table 5. Refresh interval (in iterations) and a differentiable-risk variant on NoisyMNIST, ; mean and are measured at the stopping epoch (PDF: Table R7).
| (default) | Differentiable risk | ||||
|---|---|---|---|---|---|
| ACC (%) | 93.30 | 93.12 | 92.58 | 91.43 | 87.96 |
| NMI (%) | 85.53 | 85.30 | 84.71 | 83.38 | 79.84 |
| Mean | 0.41 | 0.42 | 0.44 | 0.47 | 1.37 |
| Mean | 0.67 | 0.66 | 0.65 | 0.63 | 0.18 |
| Time (s/epoch) | 0.82 | 0.45 | 0.36 | 0.34 | 1.36 |
In Table 5, accuracy degrades gracefully up to and more clearly for , consistent with Remark 2. The differentiable variant inflates the risk and drives toward zero, i.e., it switches the guidance off, exactly as the argument above predicts.
Q5. How often does the projection activate? Distribution of ? Is the 1.78 gain significant?
Table 6. Projection activity per third of training on NoisyMNIST, , ten seeds; the full distributions are shown in Fig. R1 (PDF: Table 7(b)).
| First third | Middle third | Last third | |
|---|---|---|---|
| Iterations with (%) | 38.6 | 17.2 | 8.9 |
| Mean | 0.04 | 0.17 | 0.26 |
| Mean when | −0.37 | −0.31 | −0.24 |
Table 7. Per-seed ACC (%) of Step 3 (risk-scaled KL, unconstrained) and Step 4 (full RTG-IMVC) on NoisyMNIST, (PDF: Table R6).
| Seed | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | Mean ± std |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Step 3 | 91.81 | 91.54 | 91.26 | 89.69 | 93.54 | 90.71 | 92.22 | 91.31 | 91.86 | 91.27 | 91.52 ± 1.00 |
| Step 4 | 93.74 | 93.60 | 93.07 | 92.08 | 94.55 | 93.31 | 93.52 | 93.01 | 93.23 | 92.89 | 93.30 ± 0.64 |
| Δ | 1.93 | 2.06 | 1.81 | 2.39 | 1.01 | 2.60 | 1.30 | 1.70 | 1.37 | 1.62 | 1.78 ± 0.49 |
Table 6 shows that conflicts are frequent early in training and become rare as the teacher and the representation agree. In Table 7, Step 4 wins on all ten paired seeds; the paired -test gives , , the Wilcoxon signed-rank test gives , and the seed standard deviation drops from 1.00 to 0.64.
Q6 / W2. Per-iteration complexity, GPU memory, and mini-batch support.
With batch size , missing pairs per batch, transferred neighbors, embedding size , per-sample encoder cost , and parameters (App. G.7), one iteration costs so one epoch is linear in and , and quadratic in only through pairwise contrast. The probes do not re-run the encoder: they re-aggregate cached neighbor embeddings and re-run the cluster head without autograd. Peak memory is plus cached embeddings, with only a transient probe buffer. The "dual gradients" are two backward passes through the same graph; the projection adds .
Mini-batching. Propagation is one hop, so a batch of seed samples plus their transferred neighbors is an exact subgraph; we use this for .
Table 8. Time and memory versus , NoisyMNIST, (PDF: Table 8(b) and Fig. R2a,b).
| 5k | 10k | 20k | 70k (full) | |
|---|---|---|---|---|
| Time per epoch (s) | 0.82 | 1.69 | 3.41 | 12.37 |
| Peak memory (GB) | 1.94 | 2.02 | 2.09 | 2.21 |
| Epochs to stop | 171 | 176 | 183 | 196 |
| ACC (%) | 93.30 | 93.62 | 93.85 | 94.17 |
Table 9. Cost versus (NoisyMNIST, k) and (Caltech-5V) (PDF: Fig. R2c,d).
| 1 | 2 | 4 | 8 | 16 | |
|---|---|---|---|---|---|
| Time per epoch (s) | 0.47 | 0.53 | 0.64 | 0.82 | 1.17 |
| Peak memory (GB) | 1.91 | 1.91 | 1.92 | 1.94 | 1.97 |
| 2 | 3 | 4 | 5 | |
|---|---|---|---|---|
| Time per epoch (s) | 0.09 | 0.13 | 0.17 | 0.22 |
Time grows linearly with and , memory stays almost flat, and time grows mildly super-linearly in due to the pairwise contrast.
Q7. How does differ from PCGrad, CAGrad, and MGDA?
It relaxes an asymmetric special case rather than a symmetric multi-task rule (App. G.6, Table R1). With :
- : active iff , giving , i.e., one-sided PCGrad.
- : the constraint is the A-GEM projection (first-order non-ascent only).
- : a conflict up to is tolerated and only the excess is removed, giving the descent of Theorem 3.
MGDA and CAGrad treat both objectives symmetrically and need an inner solver; our rule protects only and is closed-form.
Table 10. Alternatives to the projection of Step 4, NoisyMNIST (ACC %, ten seeds); overhead per iteration relative to Step 3 (PDF: Table 7(c)).
| Update rule | Time overhead | ||
|---|---|---|---|
| None (Step 3) | 91.52 | 62.31 | +0.0% |
| MGDA | 91.07 | 61.58 | +6.1% |
| PCGrad | 92.24 | 63.02 | +2.3% |
| CAGrad | 92.51 | 63.37 | +9.8% |
| Ours, (one-sided PCGrad) | 92.92 | 63.88 | +1.2% |
| Ours, | 93.30 | 64.62 | +1.2% |
The relaxed rule beats the symmetric methods and its own case at the lowest overhead.
Q8 / W2. Does the projection protect an incorrect base objective at ? Releasing the constraint?
Its conservatism is bounded: only the update component along is modified, the orthogonal guidance is preserved (Eq. 41), and the projection is active in only 8.9% of iterations in the last third of training (Table 6). We also extended Fig. 8 at (Fig. R3).
Table 11. ACC (%) at under different protocols, ten seeds (PDF: Table 8(a) and Fig. R3).
| Protocol | NoisyMNIST | Caltech-5V |
|---|---|---|
| Epoch 500, with projection (Fig. 8) | 54.58 | 32.64 |
| Epoch 500, without projection | 51.47 | 29.93 |
| Epoch 500, periodic release (5 free epochs every 50) | 53.96 | 31.85 |
| Label-free stopping (reported in Table 2) | 64.62 | 42.80 |
| Mean stopping epoch | 142 | 97 |
| Oracle epoch (uses labels, reference only) | 65.83 | 43.91 |
The decline is larger without the projection, and periodic release does not help, so it is not caused by the projection; itself keeps fitting noisy transferred relations at high missingness. Fig. 8 was recorded with the stopping rule disabled, whereas all Table 2 results use the label-free rule (App. G.8), which ends within about one point of the oracle epoch. We added this as Limitation 4 (App. F.1).
We thank the reviewer again and are happy to clarify any remaining point.
We sincerely thank the reviewer for the careful reading and for recognizing the chain "structural uncertainty → consensus formation → optimization safety". All revisions are marked in orange in the updated PDF. Newly added tables in the main text of the PDF are Tables 6–8, and newly added tables and figures in the appendix are numbered R1, R2, …; the tables below are numbered for this response only, and each one indicates the corresponding PDF table.
W1 / Q1. How are LA,i(v) and ϵi(v) determined? Sensitivity to under-/over-estimation.
They are neither global constants nor tuned on validation data. Both are per-pair, label-free estimates recomputed at every teacher refresh, using only the K probe predictions that are already evaluated, so no extra forward pass is needed (new Algorithm 3, App. G.4):
Because R=D+2Lϵ depends on L and ϵ only through their product, scaling the correction by a factor c emulates under-estimation (c<1) or over-estimation (c>1) of either quantity; c=0 is the diameter-only risk.
Table 1. ACC (%) when the estimated correction 2Lϵ is scaled by c, η=0.3 (PDF: Table 7(a) and Table R10).
Table 2. Per-pair versus global estimation of L on NoisyMNIST, η=0.3 (PDF: Table 7(a)).
As Table 1 shows, for c∈[0.25,4] ACC changes by at most 0.69, 2.14, 1.06, and 0.86 points on the four datasets, whereas removing the correction (c=0) is clearly harmful. Over-estimation shifts weight toward observed views and weakens guidance through κi; under-estimation lets topology-sensitive experts keep more weight. Table 2 shows that per-pair estimates outperform a global constant.