acceptodds
Under review as a conference paper at ICLR 2027

Risk-Certified Topology Guidance for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) seeks shared clusters despite missing views. Existing methods transfer cross-view relations and graph-propagate them to construct high-confidence pseudo-targets, but transfer errors can produce stable yet inaccurate predictions. Sharp predictions do not certify topology reliability, and auxiliary guidance may conflict with contrastive representation learning.We propose the Risk-certified Topology Guidance framework (RTG-IMVC). For each missing sample, we construct an admissible topology set via donor substitution, degree-preserving edge perturbation, and neighborhood bootstrapping. The certificate combines prediction diameter and topology-cover radius to produce risk-normalized weights for a logarithmic opinion pool and bound each view's log-odds influence on the consensus teacher. We project the joint gradient onto a half-space preserving descent of the instance- and cluster-level contrastive objectives.Under local Lipschitz, smoothness, topology-coverage, and bounded-probability conditions, we prove teacher robustness to admissible perturbations within the certified radius and one-step decrease of the surrogate base objective for a suitable learning rate. RTG-IMVC makes high-confidence guidance testable and risk-controlled.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-view Clustering, Topology Guidanc, Risk Certification

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_aXhKOpenReview ↗
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
  1. 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).
  2. 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
  1. How are LA,i(v)L_{\mathcal{A},i}^{(v)} and ϵi(v)\epsilon_i^{(v)} 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 LL.

  2. Does Ri(v)R_i^{(v)} 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.

  3. What is the fusion operator for pi(f)p_i^{(f)} (attention, concatenation, averaging, etc.)? Please write out explicit formulas for Lins\mathcal{L}_{ins} and Lclu\mathcal{L}_{clu}.

  4. 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?

  5. How often does the projection actually activate (fraction of αt>0\alpha_t>0)? What is the distribution of cosine similarity between gFg_F and gGg_G? Is the 1.78%1.78\% gain from Step 3 to Step 4 statistically significant across ten seeds?

  6. Please provide per-iteration time complexity and GPU memory curves as functions of sample size, number of views, and candidate number KK. Does the framework support mini-batch subgraph sampling, or does it require full-batch graph processing?

  7. How does μ\mu functionally differ from the conflict-avoidance mechanisms of PCGrad, CAGrad, and MGDA? Is it a relaxation of these methods or a fundamentally different constraint?

  8. At η=0.5\eta=0.5, ACC on NoisyMNIST decreases from 62.09%62.09\% at epoch 100 to 54.58%54.58\% 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

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)L^{(v)}_{\mathcal A,i} and ϵi(v)\epsilon^{(v)}_i 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 KK probe predictions that are already evaluated, so no extra forward pass is needed (new Algorithm 3, App. G.4):

  • ϵ^i(v)\widehat\epsilon^{(v)}_i is a leave-one-out cover radius: each random probe is treated as a fresh admissible draw, and its dAd_{\mathcal A}-distance to the remaining probes is measured; we take the maximum.
  • L^A,i(v)\widehat L^{(v)}_{\mathcal A,i} is the largest finite-difference ratio ∥ℓ(a)−ℓ(b)∥∞/dA(a,b)\|\boldsymbol\ell(\mathbf a)-\boldsymbol\ell(\mathbf b)\|_\infty/d_{\mathcal A}(\mathbf a,\mathbf b) over probe pairs.

Because R=D^+2LϵR=\widehat D+2L\epsilon depends on LL and ϵ\epsilon only through their product, scaling the correction by a factor cc emulates under-estimation (c<1c<1) or over-estimation (c>1c>1) of either quantity; c=0c=0 is the diameter-only risk.

Table 1. ACC (%) when the estimated correction 2L^ϵ^2\widehat L\widehat\epsilon is scaled by cc, η=0.3\eta=0.3 (PDF: Table 7(a) and Table R10).

Datasetc=0c=0c=0.25c=0.25c=0.5c=0.5c=1c=1 (ours)c=2c=2c=4c=4
NoisyMNIST90.3692.8893.1493.3093.0992.61
Caltech-5V40.4248.3649.8750.5049.9448.71
BBC63.1267.0367.8868.0967.7167.21
Handwritten77.0580.2180.5580.7380.4479.87

Table 2. Per-pair versus global estimation of LL on NoisyMNIST, η=0.3\eta=0.3 (PDF: Table 7(a)).

EstimatorACC (%)NMI (%)
Per-pair L^\widehat L (ours)93.3085.53
Global L^\widehat L (maximum over the batch)92.4784.69
Global L^\widehat L (median over the batch)92.9585.12

As Table 1 shows, for c∈[0.25,4]c\in[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=0c=0) is clearly harmful. Over-estimation shifts weight toward observed views and weakens guidance through κi\kappa_i; under-estimation lets topology-sensitive experts keep more weight. Table 2 shows that per-pair estimates outperform a global constant.

Q2. Does Ri(v)R^{(v)}_i measure the reliability of the transferred topology itself, or the encoder's sensitivity to it?

By design, RR 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), RR 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 γ\gamma that controls how strongly the encoder uses topology (γ=0\gamma=0: topology-free; γ=1\gamma=1: RTG-IMVC). The protocol is in App. H.1.

Table 3. Risk diagnostic on NoisyMNIST, η=0.3\eta=0.3. ΔACC is the accuracy change caused by the corruption; AUROC measures how well a score detects corrupted rows (PDF: Table 6(a)).

γ\gammaACC (clean)ΔACC (ours)ΔACC (no certificate)AUROC of RRAUROC of confidenceSpearman(RR, neighbor error)
0 (topology-free)86.41−0.14−0.210.510.530.03
0.591.86−1.74−4.950.790.570.48
1 (default)93.30−1.83−6.620.870.590.62

Table 4. Risk quintiles on clean NoisyMNIST, η=0.3\eta=0.3 (PDF: Table 6(b) and Table R4).

Q1 (low RR)Q2Q3Q4Q5 (high RR)
Neighbor purity (%)94.890.383.672.955.4
Teacher accuracy (%)98.196.493.287.576.0
Weight of the transferred expert ω\omega0.290.230.180.130.09
Guidance weight κi\kappa_i0.880.730.590.470.37

Table 3 shows the two regimes. With γ=1\gamma=1, RR detects corrupted rows far better than prediction confidence, and the accuracy loss is limited. With γ=0\gamma=0, R≈0R\approx0, 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 pi(f)\mathbf p^{(f)}_i and explicit Lins\mathcal L_{\mathrm{ins}}, Lclu\mathcal L_{\mathrm{clu}}.

The fused prediction uses instance-level attention fusion over all view embeddings of sample ii (observed or transferred) (App. G.2, Eq. 45):

πi(v)=exp⁡(w⊤tanh⁡(Wahi(v)))∑uexp⁡(w⊤tanh⁡(Wahi(u))),hi(f)=∑vπi(v)hi(v),pi(f)=softmax(g(f)(hi(f))/Tp).\pi_i^{(v)}=\frac{\exp(\mathbf w^\top\tanh(\mathbf W_a\mathbf h_i^{(v)}))}{\sum_{u}\exp(\mathbf w^\top\tanh(\mathbf W_a\mathbf h_i^{(u)}))},\quad \mathbf h_i^{(f)}=\sum_v\pi_i^{(v)}\mathbf h_i^{(v)},\quad \mathbf p_i^{(f)}=\mathrm{softmax}\big(g^{(f)}(\mathbf h_i^{(f)})/T_p\big).

With batch B\mathcal B, cosine similarity ss, cluster-assignment columns qc(v)\mathbf q_c^{(v)}, and Ω=V(V−1)\Omega=V(V-1) (App. G.3, Eq. 46):

Lins=−1∣B∣Ω∑v≠w∑i∈Blog⁡es(hi(v),hi(w))/Tins∑j∈Bes(hi(v),hj(w))/Tins,\mathcal L_{\mathrm{ins}}=-\frac{1}{|\mathcal B|\Omega}\sum_{v\neq w}\sum_{i\in\mathcal B}\log\frac{e^{s(\mathbf h_i^{(v)},\mathbf h_i^{(w)})/T_{\mathrm{ins}}}}{\sum_{j\in\mathcal B}e^{s(\mathbf h_i^{(v)},\mathbf h_j^{(w)})/T_{\mathrm{ins}}}},

Lclu=−1CΩ∑v≠w∑c=1Clog⁡es(qc(v),qc(w))/Tclu∑kes(qc(v),qk(w))/Tclu−∑vH(qˉ(v)).\mathcal L_{\mathrm{clu}}=-\frac{1}{C\Omega}\sum_{v\neq w}\sum_{c=1}^{C}\log\frac{e^{s(\mathbf q_c^{(v)},\mathbf q_c^{(w)})/T_{\mathrm{clu}}}}{\sum_{k}e^{s(\mathbf q_c^{(v)},\mathbf q_k^{(w)})/T_{\mathrm{clu}}}}-\sum_{v}H(\bar{\mathbf q}^{(v)}).

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 ∑sωi(s)R‾i(s)=∣Pi∣/Zi−εR\sum_s\omega_i^{(s)}\overline R_i^{(s)}=|\mathcal P_i|/Z_i-\varepsilon_R and ZiZ_i decreases in every risk, both κi\kappa_i and λt\lambda_t 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 2Lθηt∥dt⋆∥22L_\theta\eta_t\|\mathbf d_t^\star\|_2, so with per-iteration refresh the staleness is first order in the step size.

Table 5. Refresh interval rr (in iterations) and a differentiable-risk variant on NoisyMNIST, η=0.3\eta=0.3; mean RR and κi\kappa_i are measured at the stopping epoch (PDF: Table R7).

r=1r=1 (default)r=5r=5r=20r=20r=100r=100Differentiable risk
ACC (%)93.3093.1292.5891.4387.96
NMI (%)85.5385.3084.7183.3879.84
Mean RR0.410.420.440.471.37
Mean κi\kappa_i0.670.660.650.630.18
Time (s/epoch)0.820.450.360.341.36

In Table 5, accuracy degrades gracefully up to r=5r=5 and more clearly for r≥20r\geq20, consistent with Remark 2. The differentiable variant inflates the risk and drives κi\kappa_i toward zero, i.e., it switches the guidance off, exactly as the argument above predicts.


Q5. How often does the projection activate? Distribution of cos⁡(gF,gG)\cos(\mathbf g_F,\mathbf g_G)? Is the 1.78 gain significant?

Table 6. Projection activity per third of training on NoisyMNIST, η=0.3\eta=0.3, ten seeds; the full distributions are shown in Fig. R1 (PDF: Table 7(b)).

First thirdMiddle thirdLast third
Iterations with αt>0\alpha_t>0 (%)38.617.28.9
Mean cos⁡(gF,gG)\cos(\mathbf g_F,\mathbf g_G)0.040.170.26
Mean cos⁡\cos when αt>0\alpha_t>0−0.37−0.31−0.24

Table 7. Per-seed ACC (%) of Step 3 (risk-scaled KL, unconstrained) and Step 4 (full RTG-IMVC) on NoisyMNIST, η=0.3\eta=0.3 (PDF: Table R6).

Seed12345678910Mean ± std
Step 391.8191.5491.2689.6993.5490.7192.2291.3191.8691.2791.52 ± 1.00
Step 493.7493.6093.0792.0894.5593.3193.5293.0193.2392.8993.30 ± 0.64
Δ1.932.061.812.391.012.601.301.701.371.621.78 ± 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 tt-test gives t9=11.46t_9=11.46, p=1.1×10−6p=1.1\times10^{-6}, the Wilcoxon signed-rank test gives p=0.002p=0.002, 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 BB, MB≈ηBVM_B\approx\eta BV missing pairs per batch, kk transferred neighbors, embedding size DD, per-sample encoder cost cencc_{\mathrm{enc}}, and PP parameters (App. G.7), one iteration costs O(BVcenc+V2B2D+MBK(kD+DC)+P),\mathcal O\big(BVc_{\mathrm{enc}}+V^2B^2D+M_BK(kD+DC)+P\big), so one epoch is linear in NN and KK, and quadratic in VV only through pairwise contrast. The KK probes do not re-run the encoder: they re-aggregate cached neighbor embeddings and re-run the cluster head without autograd. Peak memory is O(P+BVD+V2B2)\mathcal O(P+BVD+V^2B^2) plus cached embeddings, with only a transient O(MBKC)\mathcal O(M_BKC) probe buffer. The "dual gradients" are two backward passes through the same graph; the projection adds O(P)\mathcal O(P).

Mini-batching. Propagation is one hop, so a batch of seed samples plus their kk transferred neighbors is an exact subgraph; we use this for N>5,000N>5{,}000.

Table 8. Time and memory versus NN, NoisyMNIST, η=0.3\eta=0.3 (PDF: Table 8(b) and Fig. R2a,b).

NN5k10k20k70k (full)
Time per epoch (s)0.821.693.4112.37
Peak memory (GB)1.942.022.092.21
Epochs to stop171176183196
ACC (%)93.3093.6293.8594.17

Table 9. Cost versus KK (NoisyMNIST, N=5N=5k) and VV (Caltech-5V) (PDF: Fig. R2c,d).

KK124816
Time per epoch (s)0.470.530.640.821.17
Peak memory (GB)1.911.911.921.941.97
VV2345
Time per epoch (s)0.090.130.170.22

Time grows linearly with NN and KK, memory stays almost flat, and time grows mildly super-linearly in VV due to the V2V^2 pairwise contrast.


Q7. How does μ\mu 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 d0=gF+λgG\mathbf d_0=\mathbf g_F+\lambda\mathbf g_G:

  • μ=1\mu=1: active iff ⟨gF,gG⟩<0\langle\mathbf g_F,\mathbf g_G\rangle<0, giving d⋆=gF+λ(gG−⟨gF,gG⟩∥gF∥2gF)\mathbf d^\star=\mathbf g_F+\lambda\big(\mathbf g_G-\tfrac{\langle\mathbf g_F,\mathbf g_G\rangle}{\|\mathbf g_F\|^2}\mathbf g_F\big), i.e., one-sided PCGrad.
  • μ→0\mu\to0: the constraint ⟨gF,d⟩≥0\langle\mathbf g_F,\mathbf d\rangle\ge0 is the A-GEM projection (first-order non-ascent only).
  • 0<μ<10<\mu<1: a conflict up to (1−μ)∥gF∥2(1-\mu)\|\mathbf g_F\|^2 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 FF 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 ruleη=0.3\eta=0.3η=0.5\eta=0.5Time overhead
None (Step 3)91.5262.31+0.0%
MGDA91.0761.58+6.1%
PCGrad92.2463.02+2.3%
CAGrad92.5163.37+9.8%
Ours, μ=1\mu=1 (one-sided PCGrad)92.9263.88+1.2%
Ours, μ=0.4\mu=0.493.3064.62+1.2%

The relaxed rule beats the symmetric methods and its own μ=1\mu=1 case at the lowest overhead.


Q8 / W2. Does the projection protect an incorrect base objective at η=0.5\eta=0.5? Releasing the constraint?

Its conservatism is bounded: only the update component along gF\mathbf g_F 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 η=0.5\eta=0.5 (Fig. R3).

Table 11. ACC (%) at η=0.5\eta=0.5 under different protocols, ten seeds (PDF: Table 8(a) and Fig. R3).

ProtocolNoisyMNISTCaltech-5V
Epoch 500, with projection (Fig. 8)54.5832.64
Epoch 500, without projection51.4729.93
Epoch 500, periodic release (5 free epochs every 50)53.9631.85
Label-free stopping (reported in Table 2)64.6242.80
Mean stopping epoch14297
Oracle epoch (uses labels, reference only)65.8343.91

The decline is larger without the projection, and periodic release does not help, so it is not caused by the projection; FF 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.

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

This paper addresses unreliable cross-view relation transfer and optimization conflicts in Incomplete Multi-View Clustering (IMVC) by proposing the Risk-Certified Topology Guidance (RTG-IMVC) framework. For each missing sample-view pair, RTG-IMVC constructs perturbed candidate topologies to estimate a sample-level transfer-risk score based on prediction diameter and a topology-cover radius. These risk scores weight view-specific and fused predictors in a logarithmic opinion pool (LogOP) to form a consensus teacher, while a half-space gradient projection ensures that auxiliary guidance does not interfere with the descent of the primary contrastive objectives. Experiments across eight multi-view benchmarks are conducted to evaluate clustering performance under varying missing rates.

Critical strengths
  1. Well-Motivated Problem: Addressing the miscalibration of high-confidence predictions under noisy transferred graphs is well-motivated and appropriately situated within the imputation-free IMVC and safe-learning literature.

  2. Principled Integration of Weighting and Gradient Projection: Combining inverse-risk logarithmic pooling (Theorem 2) with a minimum-change half-space gradient projection (Theorem 3) provides a sensible mechanism to down-weight unstable views while protecting the primary representation learning objective.

  3. Broad Empirical Scope and Plug-and-Play Potential: The method is evaluated across eight datasets against seven recent baselines, and Appendix D.2 shows that the risk-weighted guidance module can be integrated into existing frameworks (DSIMVC, DVIMC, DDR-IMVC) to improve performance.

Critical weaknesses
  1. Unclear Presentation and Undefined Backbone Network: The paper is not self-contained and leaves core modeling components undefined. Section 3.1 (Line 193) states that relation transfer produces a normalized adjacency row a~i(v)\tilde{a}_i^{(v)}, and Line 197 states that "the network also outputs the nominal view prediction pi(v)p_i^{(v)} and fused prediction pi(f)p_i^{(f)}". However, the paper never defines what "the network" actually is, in either the main text or the appendix. This includes the mathematical formulation of the initial relation transfer a~i(v)\tilde{a}_i^{(v)}, the graph encoder architecture, and the feature fusion mechanism.

  2. Abrupt and Unexplained Terminology ("Expert" and "Consensus Teacher"): Key concepts are used without prior introduction. In Section 3.3 (Line 242), the text abruptly introduces "the risk of expert ss" over Pi={1,…,V,f}\mathcal{P}_i = \lbrace 1, \dots, V, f \rbrace without previously defining what an "expert" represents in the architecture or how the view-specific experts and consensus teacher are parameterized.

  3. Ambiguous Definition of Transfer-Risk and Overstated Certification Claims: While Eq. (6) defines transfer risk as Ri(v)=D^i(v)+2LA,i(v)ϵi(v)R_i^{(v)} = \hat{D}_i^{(v)} + 2 L_{\mathcal{A},i}^{(v)} \epsilon_i^{(v)}, the paper never explains how the local Lipschitz constant LA,i(v)L_{\mathcal{A},i}^{(v)} and topology-cover radius ϵi(v)\epsilon_i^{(v)} are numerically computed in practice (including in Algorithm 1 and the "Cover-only risk" ablation in Table 3). Moreover, as conceded in Section 3.2 (Line 237) and Appendix F.1 (Line 1412), without certified upper bounds on these quantities, Ri(v)R_i^{(v)} is merely an "empirical sensitivity score" rather than a formal risk certificate, which overstates the paper's theoretical claims.

  4. Lack of Reproducibility (Missing Contrastive Loss Formulations): In Section 3.1 (Eq. 3, Line 199), the protected objective is defined as F(θ)=Lins(θ)+Lclu(θ)F(\theta) = \mathcal{L}_{\text{ins}}(\theta) + \mathcal{L}_{\text{clu}}(\theta), denoting the instance- and cluster-level contrastive losses. However, the explicit mathematical formulations for Lins\mathcal{L}_{\text{ins}} and Lclu\mathcal{L}_{\text{clu}} are completely omitted from both the main paper and the appendix, preventing reproduction of the model and verification of the LFL_F-smoothness assumption in Theorem 3.

  5. Imprecise Terminology ("Missing Views" vs. "Missing Sample-View Pairs"): The manuscript frequently conflates missing views with missing sample-view instances. For example, the Abstract (Line 015) refers to constructing a topology set "for each missing sample" and the text repeatedly writes "missing views", whereas Eq. (1) actually formulates partially observed samples with missing sample-view pairs (i,v)(i, v) where Oi≠∅\mathcal{O}_i \neq \emptyset.

  6. Inconsistent Task Terminology in Figure 1 (Classification vs. Clustering): Figure 1(c) (Lines 061, 072) reports "Classification Error Rate (%)", whereas the paper addresses unsupervised clustering and evaluates ACC, NMI, and ARI in Section 4. Labeling unsupervised clustering error as classification error is inaccurate and inconsistent with the task definition.

  7. Small Dataset Scale and Questionable Checkpoint Selection: All eight benchmark datasets are small (up to 5,000 samples), with larger datasets like NoisyMNIST and 3V_Fashion_MV explicitly subsampled to 5,000 samples (Line 328). Because candidate topology generation (Eq. 4) perturbs NN-dimensional adjacency vectors per missing sample-view pair, scalability to large datasets remains untested. Furthermore, Appendix D.5 (Lines 1236–1241, Fig. 8) reveals that at η=0.5\eta = 0.5, fixed-epoch ACC drops by 7.51% on NoisyMNIST and 10.16% on Caltech-5V between epochs 100 and 500, noting that "These fixed-epoch measurements should be distinguished from selected-checkpoint results" in Table 2. Reporting peak selected-checkpoint results when unsupervised training suffers from severe late-epoch degradation raises concerns regarding evaluation rigor and how checkpoints were selected without ground-truth labels.

Questions for authors
  1. Missing Formulations: What are the exact mathematical formulations for (a) the instance- and cluster-level contrastive losses Lins\mathcal{L}_{\text{ins}} and Lclu\mathcal{L}_{\text{clu}} in Eq. (3), (b) the backbone network and fusion module producing pi(v)p_i^{(v)} and pi(f)p_i^{(f)}, and (c) the nominal relation transfer a~i(v)\tilde{a}_i^{(v)}?

  2. Calculation of Certificate Terms: How are the Lipschitz constant LA,i(v)L_{\mathcal{A},i}^{(v)} and cover radius ϵi(v)\epsilon_i^{(v)} in Eq. (6) quantitatively calculated during training and in the "Cover-only risk" ablation in Table 3?

  3. Checkpoint Selection in Unsupervised Clustering: Given the 7.5% to 10.2% ACC degradation between epochs 100 and 500 shown in Fig. 8 at η=0.5\eta = 0.5, how were the "selected-checkpoint results" in Table 2 chosen without access to ground-truth labels, and why did the stopping criterion ϵstop\epsilon_{\text{stop}} in Section 3.5 not halt training before this drop?

  4. Scalability and Computational Cost: How does RTG-IMVC scale to full datasets (e.g., full 70,000-sample NoisyMNIST instead of the 5,000-sample subset)? Additionally, since topology probing takes 58.5% of the per-iteration runtime (Table 8), why does RTG-IMVC have lower total FLOPs (6.83G vs. 8.74G) and memory usage than single-pass ICMVC in Table 7?

7 comments

We thank the reviewer for the very concrete comments, which have substantially improved the clarity of the paper. All revisions are marked in orange in the updated PDF, and a new appendix (App. G) makes the paper self-contained. 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(b,c). Undefined backbone network, fusion module, and nominal relation transfer.

We now specify every component (Sec. 3.1 and App. G.2):

  • Relation transfer (Eq. 43). For a missing pair (i,v)(i,v) and an observed donor view u∈Oiu\in\mathcal O_i, let Nk(u→v)(i)\mathcal N_k^{(u\to v)}(i) be the kk nearest neighbors of ii in view uu among samples observing both uu and vv. The donor row is a~ij(v←u)∝es(xi(u),xj(u))/σ,1[j∈Nk(u→v)(i)]\widetilde a^{(v\leftarrow u)}_{ij}\propto e^{s(\mathbf x_i^{(u)},\mathbf x_j^{(u)})/\sigma}\\,\mathbb 1[j\in\mathcal N_k^{(u\to v)}(i)], normalized to sum to one, and the nominal row is a~i(v)=1∣Oi∣∑u∈Oia~i(v←u)\widetilde{\mathbf a}^{(v)}_i=\frac{1}{|\mathcal O_i|}\sum_{u\in\mathcal O_i}\widetilde{\mathbf a}^{(v\leftarrow u)}_i. The donor candidates in Eq. (4) are exactly these rows.
  • Backbone (Eq. 44). View-specific encoders fθ(v)f^{(v)}_\theta and cluster heads gθ(v)g^{(v)}_\theta. For an observed pair, hi(v)=fθ(v)(xi(v))\mathbf h^{(v)}_i=f^{(v)}_\theta(\mathbf x^{(v)}_i); for a missing pair and any admissible row a\mathbf a, a one-hop propagation gives hi(v)(a)=∑jajfθ(v)(xj(v))\mathbf h^{(v)}_i(\mathbf a)=\sum_j a_j f^{(v)}_\theta(\mathbf x^{(v)}_j) and zi,θ(v)(a)=gθ(v)(hi(v)(a))\mathbf z^{(v)}_{i,\theta}(\mathbf a)=g^{(v)}_\theta(\mathbf h^{(v)}_i(\mathbf a)). "The network" in Sec. 3.1 refers to this model.
  • Fusion (Eq. 45). Instance-level attention πi(v)=softmaxv(w⊤tanh⁡(Wahi(v)))\pi^{(v)}_i=\mathrm{softmax}_v\big(\mathbf w^\top\tanh(\mathbf W_a\mathbf h^{(v)}_i)\big), hi(f)=∑vπi(v)hi(v)\mathbf h^{(f)}_i=\sum_v\pi^{(v)}_i\mathbf h^{(v)}_i, and pi(f)=softmax(g(f)(hi(f))/Tp)\mathbf p^{(f)}_i=\mathrm{softmax}(g^{(f)}(\mathbf h^{(f)}_i)/T_p).

W2. Abrupt terminology ("expert", "consensus teacher").

Both terms are now defined before first use (Sec. 3.3) and in a terminology list (App. G.1). An expert is a prediction head whose output enters the consensus: an observed-view prediction, a transferred prediction of a missing pair, or the fused prediction. The consensus teacher is their risk-weighted logarithmic opinion pool Qi\mathbf Q_i, which is detached and used as the target of the fused prediction.

W3 / Q2. How are LA,i(v)L^{(v)}_{\mathcal A,i} and ϵi(v)\epsilon^{(v)}_i computed (training, Algorithm 1, "Cover-only risk")? Overstated certification claims.

Computation. Both are per-pair, label-free estimates recomputed whenever the teacher is refreshed, using only the KK probe predictions that are already evaluated (new Algorithm 3, App. G.4; Algorithm 1 now calls it):

  1. D^\widehat D as in Eq. (6);
  2. ϵ^=max⁡b∈Srandmin⁡b′∈S∖bdA(b,b′)\widehat\epsilon=\max_{\mathbf b\in\mathcal S_{\mathrm{rand}}}\min_{\mathbf b'\in\mathcal S\setminus\\{\mathbf b\\}}d_{\mathcal A}(\mathbf b,\mathbf b'), the leave-one-out cover radius of the random probes;
  3. L^=max⁡b≠b′∣ℓ(b)−ℓ(b′)∣∞/dA(b,b′)\widehat L=\max_{\mathbf b\neq\mathbf b'}\\|\boldsymbol\ell(\mathbf b)-\boldsymbol\ell(\mathbf b')\\|_\infty/d_{\mathcal A}(\mathbf b,\mathbf b');
  4. R=D^+2L^ϵ^R=\widehat D+2\widehat L\widehat\epsilon.

The ablations in Table 3 of the paper use the same estimators: Diameter-only sets R=D^R=\widehat D, and Cover-only sets R=2L^ϵ^R=2\widehat L\widehat\epsilon. The cost is O(K2k+KC)\mathcal O(K^2k+KC) per missing pair.

Certification claims. We agree with the reviewer and revised the wording. The abstract now states that the certificate is valid under the stated coverage and Lipschitz assumptions, and "strictly risk-controlled" was replaced by "risk-controlled under these assumptions". Sec. 3.2 states that, with estimated LL and ϵ\epsilon, RR is operationally an empirical sensitivity score. It is an exact certificate only when the admissible set equals the probe set (ϵ=0\epsilon=0, App. C.1). The quality of the admissible set is controlled explicitly (Remark 1, App. G.4): every probe stays within a bounded dAd_{\mathcal A}-radius of a generator row, namely ∑j∈Raj≈ρ\sum_{j\in\mathcal R}a_j\approx\rho for edge replacement and at most β\beta for bootstrap.

To make the operational meaning concrete, we corrupt 30% of the transferred rows on NoisyMNIST by replacing half of their neighbors with other-class samples (labels are used only for this diagnostic) and vary the propagation weight γ\gamma (γ=0\gamma=0: the encoder ignores topology; γ=1\gamma=1: RTG-IMVC).

Table 1. Risk diagnostic on NoisyMNIST, η=0.3\eta=0.3. ΔACC is the accuracy change caused by the corruption; AUROC measures how well a score detects corrupted rows (PDF: Table 6(a)).

γ\gammaACC (clean)ΔACC (ours)ΔACC (no certificate)AUROC of RRAUROC of confidenceSpearman(RR, neighbor error)
0 (topology-free)86.41−0.14−0.210.510.530.03
0.591.86−1.74−4.950.790.570.48
1 (default)93.30−1.83−6.620.870.590.62

Table 2. ACC (%) when the estimated correction 2L^ϵ^2\widehat L\widehat\epsilon is scaled by cc to emulate under-/over-estimation, η=0.3\eta=0.3; c=0c=0 is the diameter-only risk (PDF: Table 7(a) and Table R10).

Datasetc=0c=0c=0.25c=0.25c=0.5c=0.5c=1c=1 (ours)c=2c=2c=4c=4
NoisyMNIST90.3692.8893.1493.3093.0992.61
Caltech-5V40.4248.3649.8750.5049.9448.71

Table 1 shows that, as a detector of corrupted rows, RR is far stronger than prediction confidence, and that it correctly stays near zero when the encoder ignores topology. Table 2 shows that the score is robust to moderate mis-estimation of LL and ϵ\epsilon, while removing the correction term is harmful. On Caltech-5V, the c=0c=0 value equals the Diameter-only result in Table 3 of the paper.


W4 / Q1(a). Missing contrastive loss formulations; LFL_F-smoothness.

The explicit formulas are now Eq. (46) (App. G.3), referenced from Sec. 3.1. With batch B\mathcal B, cosine similarity ss, and Ω=V(V−1)\Omega=V(V-1):

Lins=−1∣B∣Ω∑v≠w∑ilog⁡es(hi(v),hi(w))/Tins∑jes(hi(v),hj(w))/Tins,\mathcal L_{\mathrm{ins}}=-\frac{1}{|\mathcal B|\Omega}\sum_{v\neq w}\sum_{i}\log\frac{e^{s(\mathbf h_i^{(v)},\mathbf h_i^{(w)})/T_{\mathrm{ins}}}}{\sum_{j}e^{s(\mathbf h_i^{(v)},\mathbf h_j^{(w)})/T_{\mathrm{ins}}}},

Lclu=−1CΩ∑v≠w∑clog⁡es(qc(v),qc(w))/Tclu∑kes(qc(v),qk(w))/Tclu−∑vH(qˉ(v)),\mathcal L_{\mathrm{clu}}=-\frac{1}{C\Omega}\sum_{v\neq w}\sum_{c}\log\frac{e^{s(\mathbf q_c^{(v)},\mathbf q_c^{(w)})/T_{\mathrm{clu}}}}{\sum_{k}e^{s(\mathbf q_c^{(v)},\mathbf q_k^{(w)})/T_{\mathrm{clu}}}}-\sum_v H(\bar{\mathbf q}^{(v)}),

where qc(v)\mathbf q_c^{(v)} is the cc-th column of the batch assignment matrix and HH is the entropy. Regarding Theorem 3: FF is a composition of the encoders, ℓ2\ell_2 normalization (smooth away from zero), softmax, logarithm, and entropy, so ∇F\nabla F is locally Lipschitz for smooth activations. For piecewise-linear activations the assumption holds within each linear region, which is why Theorem 3 is stated as a local, one-step result (App. G.3).

W5. "Missing views" vs. "missing sample–view pairs".

We agree. The abstract now says "for each missing sample–view pair", and Sec. 3.1 defines a missing pair (i,v)(i,v) with v∈Miv\in\mathcal M_i, noting that a sample can have several missing pairs. All topology operations act on missing pairs.


W6. "Classification error rate" in Fig. 1(c).

Thank you for spotting this. Fig. 1(c) and its caption now read "Clustering Error Rate", defined as 1−ACC1-\mathrm{ACC} after Hungarian matching.


W7 / Q3. Checkpoint selection without labels; why did εstop\varepsilon_{\mathrm{stop}} not halt training before the drop in Fig. 8?

No ground-truth labels are used for model selection (App. G.8). After every epoch we compute the relative change of the training objective LRTG\mathcal L_{\mathrm{RTG}} over a window of ww epochs and stop when it falls below εstop\varepsilon_{\mathrm{stop}}; the model at that epoch is evaluated. ww and εstop\varepsilon_{\mathrm{stop}} are shared by all datasets, missing rates, and seeds, and all results in Table 2 of the paper are produced this way. The curves in Fig. 8 were recorded with the stopping rule disabled, precisely to expose long-horizon behavior, which is why they continue to epoch 500. We now say this explicitly in Sec. 3.5, App. D.5, and App. G.8.

Table 3. ACC (%) at η=0.5\eta=0.5 under different training and model-selection protocols, ten seeds (PDF: Table 8(a) and Fig. R3).

ProtocolNoisyMNISTCaltech-5V
Fixed epoch 500 (Fig. 8)54.5832.64
Fixed epoch 500, without projection51.4729.93
Label-free stopping (reported in Table 2)64.6242.80
Mean stopping epoch14297
Oracle epoch (uses labels, reference only)65.8343.91

Table 3 shows that the label-free rule ends within 1.21 (NoisyMNIST) and 1.11 (Caltech-5V) points of the oracle epoch. The late decline is larger without the projection, so it stems from the base contrastive objective fitting noisy transferred relations at high missingness rather than from our optimization (see also our answer to Reviewer aXhK, Q8).

Reviewer_HR6nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses a classic task in clustering, termed Incomplete Multi-View Clustering (IMVC), which leverages cross-view associations to recover the shared cluster structures from partially observed multi-view data. Unreliable transferred topologies and guidance conflicts in consistent representation learning are two key issues focused by this submission. The proposed method RTG-IMVC uses four key modules to alleviate the aforementioned questions, where (1) Topology Construction uses donor substitution and edge replacement to probe sensitivity; (2) Risk Certificate combines prediction diameter and topology-cover radius to quantify missing-view transfer risk; (3) Risk-Normalized LogOP uses inverse-risk weights to suppress unreliable experts and scale guidance; and (4) Base-Safe Optimization projects gradients into a half-space to guarantee the descent of the protected contrastive objective while retaining guidance. Experiments on eight benchmarks show that RTG-IMVC achieves the best performance in 68 of 72 settings, outperforming 7 baselines by up to 8.34% in accuracy, and the effectiveness of the proposed modules has been verified in improving the performance of model by a large margin of over 10%. The runtime analysis confirms it is the fastest method per iteration, demonstrating high efficiency and robustness.

Critical strengths
  1. The proposed method uses methods of substitute donors and certified robustness to alleviate the unreliable transferred topologies; and it designs a half-space to preserve the base features. Such procedures are reasonable, and may be an alternative to current latent feature learning by neural nets.
  2. Mathematical proofs are given for the controllable algorithmic process, including the certified topology stability for risk assessment, the bounded expert influence for risk control, and the One-step base-objective descent for the convergence of the optimization process.
Critical weaknesses
  1. The motivation is not strong. There have been a dozen of IMVC paper using neural feature learner. The raised challenges are either narrow or general, i.e. the unreliable transferred topologies are not applicable to the neural approaches and the consistent representation learning between views is the root question for multi-modal learning. It may not be the truth that the proposed method addresses all challenges and does not introduce new issues in IMVC and may beat all the baselines. A justification of this solution as compared to the sota IMVC solutions is needed.
  2. The time complexity should be high, as its estimation process may introduce loop operators. The number of features is not reported. However, as in Fig. 6, over a second is required to run for a single iteration on 5000 samples. So it may not be scalable. Moreover, case studies for the working mechanism of the proposed modules are missing.
Questions for authors
  1. Please provide a time complexity analysis, and report the time cost between RTG-IMVC and a sota deep IMVC algorithm when converging.
  2. How to control the quality of the admissible topology, and what is the underlying principle that the transfer-risk certificate works?
  3. How is the no. of iteration changes along with the increase in the no. of datasets and classes?
3 comments

We thank the reviewer for the thoughtful comments on motivation and efficiency. 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. The motivation is not strong; are unreliable transferred topologies relevant to neural approaches? Justification against SOTA IMVC solutions.

We respectfully clarify that the problem we address arises inside neural IMVC models. Recent neural methods such as ICMVC (AAAI'24) and GHICMC (AAAI'25) infer missing-view representations by propagating over relations transferred from observed views, and they then use high-confidence predictions as guiding targets. A wrong transferred neighbor therefore enters the learned features directly, while the confidence of the resulting target need not change. We do not claim to solve cross-view consistency learning in general. Our contribution is a reliability layer for this specific, measurable failure mode, and the introduction now states this explicitly. We support it with three pieces of evidence.

(1) The failure mode exists in neural methods. We replace a fraction ζ\zeta of every transferred neighborhood with other-class samples and retrain each method.

Table 1. Effect of transferred-topology errors on neural methods, NoisyMNIST, η=0.3\eta=0.3. "Target confidence" is the mean maximum probability of ICMVC's high-confidence guiding targets (PDF: Table R3).

Corrupted fraction ζ\zeta0%10%20%30%40%
ICMVC, ACC (%)76.0172.8468.9364.1758.66
ICMVC, target confidence0.930.930.920.920.91
GHICMC, ACC (%)78.6275.3171.2066.5861.04
RTG-IMVC, ACC (%)93.3092.7191.9891.1289.85
RTG-IMVC, mean risk RR0.410.520.660.810.97

Table 1 shows that ICMVC and GHICMC lose more than 17 ACC points, while the confidence of ICMVC's guiding targets hardly moves: confidence does not reveal transferred-topology errors. The mean risk of RTG-IMVC grows with ζ\zeta, and its accuracy drops by only 3.45 points.

(2) The mechanism transfers to other neural IMVC methods. Adding our risk-aware guidance to DSIMVC, DVIMC, and DDR-IMVC improves ACC by 9.94, 9.76, and 8.76 points, respectively (App. D.2, Table 12 of the paper).

(3) Comparison with SOTA. RTG-IMVC is compared with seven recent methods, including GHICMC (2025), SIIHPC (2025), HSACC (2026), and DDR-IMVC (2026), and achieves the best mean in 68 of 72 settings (Table 2 of the paper). Table 4 below adds the full 70k NoisyMNIST.


Q2. How is the quality of the admissible topology controlled, and what is the underlying principle of the transfer-risk certificate?

Quality control (App. G.4). The admissible set Ui(v)\mathcal U_i^{(v)} contains the rows reachable from a generator row (the nominal row, or a donor row transferred from another observed view) by one bounded operation. Remark 1 shows that these operations have explicit radii: edge replacement moves a row by dA=∑j∈Raj≈ρd_{\mathcal A}=\sum_{j\in\mathcal R}a_j\approx\rho, and bootstrap moves it by at most β\beta. Hence ρmax⁡\rho_{\max} and βmax⁡\beta_{\max} directly bound how far admissible topologies can depart from what the observed views support, and the estimated cover radius ϵ^\widehat\epsilon (Algorithm 3) measures how well the finite probes cover this set.

Principle. The certificate links two guarantees. Theorem 1: if the probes ϵ\epsilon-cover Ui(v)\mathcal U_i^{(v)} and the log-prediction is LL-Lipschitz there, then no admissible topology can move the expert's log-prediction by more than R=D^+2LϵR=\widehat D+2L\epsilon (the probe diameter plus twice the cover gap). Theorem 2: inverse-risk weighting then bounds how far any such expert can move the teacher's log-odds. Thus RR certifies the stability of a decision under plausible topologies. We test this by corrupting 30% of transferred rows and varying the weight γ\gamma with which the encoder uses topology.

Table 2. Risk diagnostic on NoisyMNIST, η=0.3\eta=0.3. ΔACC: accuracy change under corruption; AUROC: detection of corrupted rows (PDF: Table 6(a)).

γ\gammaΔACC (ours)ΔACC (no certificate)AUROC of RRAUROC of confidenceSpearman(RR, neighbor error)
0 (topology-free)−0.14−0.210.510.530.03
0.5−1.74−4.950.790.570.48
1 (default)−1.83−6.620.870.590.62

Table 2 shows that, by default, RR detects corrupted rows far better than prediction confidence and limits the accuracy loss; when the encoder ignores topology, R≈0R\approx0 and the corruption does not affect accuracy, which is the intended behavior.

W2 / Q1. Time complexity, loop operators, number of features, time to convergence against a SOTA deep IMVC method, and case studies.

Complexity (App. G.7). With batch size BB, MB≈ηBVM_B\approx\eta BV missing pairs per batch, kk transferred neighbors, embedding size DD, per-sample encoder cost cencc_{\mathrm{enc}}, and PP parameters, one iteration costs O(BVcenc+V2B2D+MBK(kD+DC)+P).\mathcal O\big(BVc_{\mathrm{enc}}+V^2B^2D+M_BK(kD+DC)+P\big). One epoch is therefore linear in NN and KK. The topology probes do not introduce loop operators over the network. The per-sample loop in Algorithm 1 is written for readability; the implementation stacks all MBKM_BK probe rows into one sparse matrix and evaluates them in a single batched pass over cached neighbor embeddings without autograd. The kkNN graphs are built once before training. An "iteration" in Fig. 6 processes all NN samples (one epoch for N≤5,000N\le5{,}000), and RTG-IMVC is the fastest of the eight methods on all eight datasets (Fig. 6 and Fig. 12).

Table 3. Time, memory, and training length versus the number of samples NN on NoisyMNIST, η=0.3\eta=0.3; one-hop subgraph mini-batches are used for N>5N>5k (PDF: Table 8(b) and Fig. R2).

NN5k10k20k70k (full)
Time per epoch (s)0.821.693.4112.37
Peak memory (GB)1.942.022.092.21
ACC (%)93.3093.6293.8594.17

Table 4. Time to convergence and accuracy against recent deep IMVC methods on NoisyMNIST, η=0.3\eta=0.3; ACC and NMI are on the full 70k set (PDF: Table 8(c)).

MethodACC (%)NMI (%)Time to converge, 5k / 70k (min)Peak memory, 70k (GB)
ICMVC77.1269.054.7 / 61.47.82
DSIMVC83.0575.107.9 / 88.75.36
DVIMC82.2774.315.6 / 70.24.11
DDR-IMVC84.2177.6410.2 / 121.56.94
RTG-IMVC94.1786.382.3 / 40.42.21

Table 5. Per-iteration cost breakdown of RTG-IMVC on NoisyMNIST, η=0.3\eta=0.3 (PDF: Table R8).

ComponentGFLOPsTime (s)
Encoder forward and propagation/fusion/heads2.000.10
Topology probes (no autograd)0.610.48
Risk and teacher construction0.010.03
Losses, two backward passes, projection4.210.21
Total6.830.82

Tables 3–5 show that time grows linearly with NN, memory stays near 2 GB, and RTG-IMVC converges fastest at both scales. The probes account for most of the wall-clock time but little computation, because they are many small sparse gathers rather than loops over the network.

Number of features. We added the feature dimension of every view.

Table 6. Feature dimension of each view (PDF: Table R2).

DatasetVVd1/⋯/dVd_1/\cdots/d_V
NGs32000 / 2000 / 2000
NoisyMNIST2784 / 784
MSRC_v1524 / 576 / 512 / 256 / 254
3V_Fashion_MV3784 / 784 / 784
Caltech-5V540 / 254 / 928 / 512 / 1984
BBC44659 / 4633 / 4665 / 4684
Handwritten476 /216 /47 / 6
3-sources33560 / 3631 / 3068

W2 / Q1 (continued). Case studies.

Case studies (App. H.3). We trace individual NoisyMNIST samples whose second view is missing, and we summarize the mechanism over risk quintiles.

Table 7. Case study on NoisyMNIST (η=0.3\eta=0.3, view 2 missing). ω\omega lists the weights of the observed, transferred, and fused experts; "Transferred" is the top cluster of the transferred expert (PDF: Table R5).

SampleLabelNeighbor purityR(2)R^{(2)} / R(f)R^{(f)}ω\omega (obs. / transf. / fused)TransferredTeacher (prob.)κi\kappa_i
#041271.000.06 / 0.030.38 / 0.29 / 0.3377 (0.97)0.90
#187330.900.18 / 0.090.45 / 0.24 / 0.3133 (0.93)0.75
#229040.700.47 / 0.220.56 / 0.17 / 0.2744 (0.81)0.58
#315790.500.93 / 0.410.66 / 0.12 / 0.2249 (0.68)0.45
#402150.301.62 / 0.700.75 / 0.08 / 0.1735 (0.57)0.37
#468820.201.88 / 0.860.78 / 0.07 / 0.1577 (0.46)0.34

Table 8. Risk-quintile analysis at η=0.3\eta=0.3 (PDF: Table 6(b) and Table R4).

DatasetStatisticQ1 (low RR)Q2Q3Q4Q5 (high RR)
NoisyMNISTNeighbor purity (%)94.890.383.672.955.4
NoisyMNISTTeacher accuracy (%)98.196.493.287.576.0
Caltech-5VNeighbor purity (%)81.272.563.051.838.6
Caltech-5VTeacher accuracy (%)78.466.955.144.731.2
BBCNeighbor purity (%)88.780.170.458.944.3
BBCTeacher accuracy (%)90.281.570.861.347.9

Table 7 shows the mechanism at work. When the transferred neighborhood is pure, the transferred expert keeps a substantial weight and the guidance is strong. As purity drops, the transferred expert starts voting for a wrong cluster, its risk rises, and its weight shrinks, so the teacher follows the observed view. In the last example both the transferred expert and the teacher are wrong, but κi=0.34\kappa_i=0.34 limits the damage of this target. Table 8 shows the same pattern on all three datasets.


Q3. How does the number of iterations change with the number of samples and classes?

Table 9. Epochs until the label-free stopping rule fires: (a) NoisyMNIST subsets of size NN (C=10C=10, η=0.3\eta=0.3); (b) the first CC digit classes of NoisyMNIST (N=5N=5k, η=0.3\eta=0.3); (c) all benchmarks (η=0.1\eta=0.1) (PDF: Table R9).

(a) NN1k2k5k10k20k70k
Epochs158164171176183196
Total time (min)0.500.962.344.9610.4040.40
(b) CC246810
Epochs92118139157171
ACC (%)99.1297.3595.4894.2193.30
(c) Dataset (NN, CC)NGs (500, 5)NoisyMNIST (5000, 10)MSRC_v1 (210, 7)3V_Fashion_MV (5000, 10)Caltech-5V (1400, 7)BBC (685, 5)Handwritten (2000, 10)3-sources (169, 6)
Epochs104168871791219814679

As Table 9 shows, the number of epochs grows only from 158 to 196 while NN grows 70-fold, so the total training time is essentially linear in NN. The number of epochs increases moderately with CC, and across the eight benchmarks it follows the same pattern.

We thank the reviewer again and hope these additions address the concerns about motivation and efficiency. We are happy to provide further details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.