acceptodds
Under review as a conference paper at ICLR 2027

When Missingness Carries Meaning: Causal Federated Multi-View Clustering with Informative View Missingness

Abstract

Federated multi-view clustering enables multiple institutions to cluster data collaboratively without sharing raw features, yet in practice many samples lack one or more views. Existing methods treat the missing pattern as a mere index: they use what is observed, impute what is not, and implicitly assume that views are missing at random. In practice, missingness often carries meaning, as which views a sample lacks is correlated with its latent semantics: clinicians typically order an imaging examination only when initial symptoms suggest severe disease, so a missing imaging view indicates a likely milder case. Ignoring this discards useful information and introduces selection bias, since each client observes only samples with an available view. We propose CFIMV, which places the missing mechanism inside a causal graph and builds three components on it: (i) a mechanism model that turns the missing pattern into additional evidence for cluster assignment and into inverse-probability weights that correct client-side selection bias; (ii) sample-level counterfactual view imputation, replacing cluster-level imputation that fills a missing view with the prototype of the currently assigned cluster and thus reinforces assignment errors; and (iii) a label-free reliability score and doubly robust prototypes that determine how much imputed views are trusted. Extensive experiments demonstrate the superiority of CFIMV against state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated learning, Multi-view clustering, Incomplete views, Counterfactual imputation, Doubly robust estimation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NU6hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of federated incomplete multi-view clustering with missing views. Unlike existing methods that typically treat missing patterns as mere data indices, this paper proposes that missing patterns themselves may contain information related to the underlying cluster structure. To address this issue, the authors introduce the CFIMC framework. Firstly, by modeling the missing mechanism, missing patterns are treated as additional clustering evidence, and inverse probability weighting is utilized to correct selection bias. Secondly, sample-level counterfactual imputation is achieved through cross-view shared semantic representations, mitigating the error self-reinforcement caused by traditional cluster-level imputation. Lastly, reliability assessment and a dual robust prototype are employed using distribution transfer weighting to reduce the impact of unreliable imputation results on clustering. The authors provide relevant theoretical analysis and conduct experimental validation on synthetic datasets and multiple real-world multi-view datasets.

Critical strengths
  1. The research perspective is novel. Unlike existing methods that merely regard missing patterns as data indices or information requiring completion, this paper further considers the potential semantic information contained within the missing patterns themselves, providing a meaningful new perspective for federated incomplete multi-view clustering.

  2. The methodological framework is complete and the logic is clear. The article explicitly points out three core issues caused by the absence of informational view, and designs corresponding solution modules for each issue. There is a relatively clear correspondence between research questions, methodological design, and theoretical analysis, which ensures good logical coherence in the overall framework.

Critical weaknesses
  1. The paper takes the structural causal model as the core theoretical foundation and proposes counterfactual view interpolation. However, from the perspective of the actual algorithm, Module II mainly learns shared semantic representations through other observable views, and then utilizes a cross-view mapping network to predict missing views. This process bears strong similarities to existing shared representation learning and cross-view prediction methods. The authors need to further clarify what capabilities causal modeling brings at the algorithmic level that traditional cross-view interpolation cannot achieve.

  2. Although the article provides a relatively comprehensive comparison of clustering performance and ablation experiments, there is still a lack of sufficient direct verification regarding the information provided by missing patterns, the correction effect of selection bias, and the estimation accuracy of the dual robust prototype.

Questions for authors
  1. According to Theorem 3.4, when Condition (P) fails, even an accurate imputation model does not guarantee consistency unless ρ=1\rho=1. However, the reliability coefficient estimated by the proposed method may be smaller than one even for an optimal imputer. Under these circumstances, in what sense is the estimator doubly robust? Can the authors provide theoretical guarantees for the actual reliability-weighted and shrinkage-based prototype estimator, particularly when pseudo-labels contain errors?

  2. Can the author further elaborate on the fundamental differences between their approach and existing cross-view imputation methods, and further explain what capabilities causal modeling brings at the algorithmic level that traditional cross-view imputation cannot achieve.

6 comments
AuthorsOpenReview ↗

W1 / Q2. What causal modeling brings beyond cross-view prediction

Thank you for the review and for pressing on what the causal view adds in practice. In response, we sharpened both the paper and the experiments.

We agree that the networks of Module II are ordinary MLPs and that its architecture resembles cross-view prediction (COMPLETER, DCP). The causal formulation does not change the network class; it changes what the networks estimate, on which population they are trained, how much they are trusted, and how they are used (new paragraph in Section 3.3):

  1. An estimand that does not shift with selection. Module II estimates the mechanism fm(c)=E[Z(m)∣C=c]f_m(c)=E[Z^{(m)}\mid C=c], which is the same for visible and missing entries (Theorem 3.2). A direct regressor estimates E[Z(m)∣Z(A),R(m)=1]E[Z^{(m)}\mid Z^{(A)},R^{(m)}=1], which differs from the corresponding regression on the missing population even with infinite data (Proposition D.10; in a two-cluster Gaussian model the gap is nonzero at every point whenever πm(1)≠πm(2)\pi_m(1)\neq\pi_m(2)). Abduction pools all visible views on the content sphere, which shrinks this gap at rate ∣Vi∣−1/2|\mathcal V_i|^{-1/2} and handles every missing pattern with one network per view.
  2. The right training population. Transport weights (1−π^)/π^(1-\hat\pi)/\hat\pi make the training risk on visible entries equal to the risk on the missing entries that the imputer actually serves (Proposition D.11), and view dropout matches their information conditions.
  3. Label-free certification. Because the mechanism is invariant, cross-fitted, transport-weighted errors on visible entries estimate the error on missing entries (Theorem 3.3). An associational imputer has no such guarantee, and its visible-entry error is optimistic when the rarely visible clusters are the hardest to impute (Theorem 3.3(ii)). This certificate is what Module III turns into fusion weights and the doubly robust correction.
  4. Coherent downstream use. The same model serves as the outcome regression of the doubly robust prototypes, whose bias we decompose exactly (Theorem 3.4 and the new Proposition D.16).

Experiment (Table 2, Appendix H.7). Modules I and III are fixed, and Module II is replaced by sample-level associational imputers with the same capacity and cross-fitting, and by our own networks trained with plain squared error. Besides ACC and NMI, the hidden views give each imputer's oracle R2R^2 on missing entries and its optimism (visible minus hidden R2R^2), the quantities that points 1–3 predict. Averages over four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean") and five seeds (η=0.5\eta=0.5); parentheses: paired change vs. full CFIMC.

Module II replaced byGranularityMNAR ACCMNAR NMIMNAR_latent ACCMNAR_latent NMIHidden-entry R²: MNAR / MNAR_latentVisible − hidden R²: MNAR / MNAR_latent
No imputation–66.58 (−1.88)61.73 (−1.06)67.08 (−1.88)60.03 (−2.28)– / –– / –
Cluster-level, FCUIFcluster65.93 (−2.54)61.58 (−1.21)66.71 (−2.25)59.79 (−2.52)0.224 / 0.4070.618 / 0.421
View meanconstant66.47 (−1.99)61.70 (−1.09)67.03 (−1.94)59.92 (−2.39)-0.240 / -0.286-0.185 / 0.081
Cross-view kNNsample66.49 (−1.97)61.43 (−1.36)70.52 (+1.56)64.13 (+1.82)-0.313 / 0.3251.231 / 0.604
Direct cross-view regressionsample68.40 (−0.07)62.63 (−0.16)69.86 (+0.90)62.53 (+0.21)-0.296 / 0.2621.177 / 0.635
Same networks, plain MSEsample68.46 (−0.00)62.97 (+0.18)69.20 (+0.23)62.10 (−0.21)-0.132 / 0.3811.082 / 0.581
Module II (ours)sample68.4762.7968.9762.31-0.233 / 0.3321.189 / 0.634
AuthorsOpenReview ↗

Q1. In what sense is the estimator doubly robust when (P) fails, and what does it guarantee with pseudo-label errors?

Your reading of Theorem 3.4 is correct: when (P) fails, an accurate imputer yields consistency only at ρ = 1, while ρ^m\hat\rho_m estimates the R2R^2 on the missing population, which is below one even for the optimal imputer whenever the view has private noise. We revised the paper as follows.

Terminology (Remark D.17). Eq. (13) is the AIPW estimating equation with the augmentation scaled by ρ. At ρ = 1 it is doubly robust in the classical sense. With ρ^m<1\hat\rho_m<1 it keeps the propensity branch, consistency under (P) for every imputer, while the imputation branch yields a quantified bias reduction rather than consistency. We now say this explicitly and corrected the sentence after Theorem 3.4.

Guarantee for the implemented estimator (new Proposition D.16). It covers the estimator we actually run: the estimated ρ^m\hat\rho_m (any limit ρˉm∈[0,1]\bar\rho_m\in[0,1], no (P) needed), the shrinkage of Eq. (14), and estimated, possibly wrong, memberships AkA_k, for any missingness mechanism.

  • (i) plim⁡μk(m)=E[Z∣Ak]+(1−ρˉm) bk(Z)+ρˉm bk(Z−g^m)\operatorname{plim}\mu^{(m)}_k=E[Z\mid A_k]+(1-\bar\rho_m)\,b_k(Z)+\bar\rho_m\,b_k(Z-\hat g_m), where bk(W)b_k(W) is the selection bias of the visible mean of WW within AkA_k.
  • (ii) ∥bk(W)∥≤sd(πW∣Ak)pkm sd(W∣Ak)≤1−pkmpkm sd(W∣Ak)\|b_k(W)\|\le\frac{\mathrm{sd}(\pi_W\mid A_k)}{p_{km}}\,\mathrm{sd}(W\mid A_k)\le\sqrt{\frac{1-p_{km}}{p_{km}}}\,\mathrm{sd}(W\mid A_k).
  • (iii) Pseudo-label errors add at most 2B P(Ak△{Y=k})/P(Ak)2B\,P(A_k\triangle\{Y=k\})/P(A_k) for bounded embeddings.
  • (iv) With a correct imputer and true memberships the bias is exactly (1−ρˉm)Bkmsel(1-\bar\rho_m)B^{\rm sel}_{km}: the visible-mean bias shrinks by the factor 1−ρˉm1-\bar\rho_m and is never amplified.

Hence, the error of the actual prototypes is controlled by the strength of selection within the estimated cluster, times (1−ρˉ)(1-\bar\rho) × the spread of the view plus ρˉ\bar\rho × the imputer residual, plus the misclassification rate.

Why ρ^m\hat\rho_m rather than ρ = 1. The label-free reliability cannot distinguish an accurate but noisy imputer from a biased one. At ρ = 1 the term Covk(π⋆,g⋆−g^m)\mathrm{Cov}_k(\pi^\star,g^\star-\hat g_m) makes a poor imputer harmful, whereas under (P), where both terms vanish, ρ only affects the variance (Theorem D.18). Empirically (η=0.5\eta=0.5, four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean"), five seeds; parentheses: paired change vs. ρ^m\hat\rho_m):

Coefficient ρ in Eq. (13)MNAR ACCMNAR NMIMNAR_latent ACCMNAR_latent NMI
ρ≡0, visible mean68.26 (−0.20)62.80 (+0.01)69.36 (+0.40)63.16 (+0.85)
ρ̂ (ours)68.4762.7968.9762.31
ρ≡1, AIPW69.63 (+1.16)63.21 (+0.43)68.52 (−0.44)61.85 (−0.46)

AIPW performs comparably / slightly better under MNAR, where the imputer is accurate, which matches Proposition D.16(iv); ρ̂ protects against poor imputers, as at η = 0.9 or on small datasets (Table 20). The last rows of the third table in our next comment compare the observed residual of the implemented estimator with the share 1−ρ^m1-\hat\rho_m predicted by Proposition D.16(iv), using oracle prototypes.

AuthorsOpenReview ↗

W2. Direct verification of the information in missing patterns, of the selection-bias correction and of the DR prototypes

Because missingness is simulated, the hidden views and the labels allow us to audit each mechanism directly (new Appendix H.9: full model, four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean"), three seeds, η=0.5\eta=0.5, server coordinates).

(a) Information in the missing patterns. I(Y;R)I(Y; R) in bits; the accuracy of the best classifier that sees the pattern alone (majority-class rate in parentheses); the error of the estimated visibility table against the truth; the share of samples whose cluster changes when the evidence Λ\Lambda is added to the feature distance, how many of them move to their true cluster, and the net change in ACC.

MechanismI(Y;R) bitsPattern-only ACC % (majority)Table MAEAssignments changed by Λ (%)Of which moved to the true cluster (%)Net ACC gain (pts)
MNAR0.94540.5 (16.2)0.0730.5332.6+0.26
MNAR_latent0.43531.5 (16.2)0.0590.1914.3-0.05

(b) Correction of the selection bias. Total variation distance between each client's visible cluster law and the population law, before and after our inverse-probability weights, and the recall of rarely visible clusters with and without Module I.

MechanismTV(P(Y | R=1), P(Y)) unweightedTV after IPW with π^\hat\piRare-cluster recall %: CFIMC / w/o mech.Common-cluster recall %: CFIMC / w/o mech.
MNAR0.1550.09976.4 / 73.684.7 / 84.7
MNAR_latent0.0920.03279.2 / 83.378.4 / 80.1

(c) Accuracy of the DR prototypes. Distance of each estimator to the oracle prototype, the mean of the hidden-view embeddings over the cluster, divided by the within-cluster standard deviation, with true and estimated memberships. The last three rows compare the observed residual of our estimator with the share 1−ρ^1-\hat\rho predicted by Proposition D.16(iv).

η = 0.5; error = ‖estimate − oracle mean of the cluster‖ / within-cluster sd, averaged over clusters and views (lower is better).

Prototype estimatorMNAR: oracle memb.MNAR: estimated memb.MNAR_latent: oracle memb.MNAR_latent: estimated memb.
Visible mean (ρ=0)0.0930.2470.1210.191
Completed-block mean (w/o DR)0.2160.2430.1480.187
Outcome regression0.2610.3040.1730.227
DR with ρ̂ (ours)0.2080.2480.1480.188
AIPW (ρ=1)0.2240.2500.1520.189
‖B_sel‖ / sd (visible-mean bias)0.0930.2470.1210.191
DR residual / ‖B_sel‖ (observed)2.0760.8671.0190.893
1 − ρ̂ (prediction of the new proposition)0.0420.0430.0290.029

We hope these additions address your concerns, and we would be glad to clarify anything further.

Reviewer_31UQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated incomplete multi-view clustering with informative view missingness. The authors propose CFIMC, which models the missing mechanism in a causal graph and uses missing patterns as additional evidence for cluster assignment. It combines inverse-probability weighting and sample-level counterfactual view imputation with doubly robust prototype estimation using reliability scores. The paper provides theoretical analyses of counterfactual identifiability and prototype estimation bias. Experiments show good clustering performance across multiple datasets and missingness settings, and ablation studies further assess the roles of individual components.

Critical strengths
  1. The motivation is clear. Missing patterns can provide evidence for cluster assignment, and the paper explicitly considers the selection bias caused by view availability.

  2. The paper is well organized and easy to follow. Each module addresses a problem raised in the introduction, and the framework diagrams help explain how the method works.

  3. The theoretical analysis explains when missing patterns provide information for clustering. The bias decomposition shows how selection bias affects prototype estimation, and how imputation error and reliability weighting enter the estimate.

  4. CFIMC performs well on multiple datasets under different missingness settings and rates. The ablation studies and parameter sensitivity analysis help assess the proposed components, while the hidden-view audit examines imputation quality.

Critical weaknesses
  1. The experiments omit FIMCFG (ICML 2025), a closely related federated incomplete multi-view clustering method discussed in the related work. The paper does not explain how HFMVC and FCUIF are adapted to this setting or which imputation method is used for the complete-view baselines. Without these details, it is difficult to judge whether the comparisons are fair. Representative centralized incomplete multi-view clustering methods would provide a useful reference if evaluated with the same missingness masks and access only to observed data.

  2. The paper does not specify how the continuous latent factor in MNAR latent is generated without labels or how it determines view observation probabilities. Without these details, the setting is difficult to reproduce, and its departure from Condition (P) is difficult to assess.

  3. The "w/o CF imputation" variant replaces sample-level counterfactual imputation with cluster-level imputation, changing both the imputation method and the prediction granularity. Lemma D.8 motivates avoiding self-confirming cluster-level imputation, but does not establish an advantage over other sample-level imputers. The ablation therefore does not isolate the benefit of the counterfactual formulation in Module II.

  4. The statement after Theorem 3.4 that accurate imputers remove selection bias is broader than the guarantee provided. With true cluster memberships and a fixed ρ\rho, Eq. (36) gives a residual bias of (1−ρ)Bsel(1 - \rho) B_{\mathrm{sel}} when the imputer is correct. The R2R^2 weight in Eq. (10) need not approach one because prediction error can remain even for a correct conditional-mean predictor. Moreover, the convergence result for the estimated reliability in Theorem 3.3 assumes Condition (P). These results do not establish that accurate imputation alone removes selection bias for the implemented estimator when (P) fails.

Questions for authors
  1. How does CFIMC compare with FIMCFG under the same experimental protocol? How is HFMVC adapted given the different client and view partitioning settings? Since the samples are assumed to be aligned, how is FCUIF's alignment module configured? Which imputation method is used for the complete-view baselines, and why was it chosen?

  2. How is the continuous latent factor in MNAR latent generated without labels, and how does it determine view observation probabilities? How is η\eta defined? On two-view datasets, retaining at least one view per sample caps the missing-block fraction at 0.5, yet η\eta reaches 0.9. How are these nominal rates implemented, and what missing fractions remain after enforcing the constraint?

  3. The caption of Table 6 refers to "three missing mechanisms," but the table reports only MNAR and MNAR latent. Was a third mechanism evaluated, and if so, what are the corresponding results? Table 2 also points to Algorithm 2 for TpreT_{\mathrm{pre}} and TlocT_{\mathrm{loc}}, but neither parameter appears in that algorithm. Where are these parameters used in the training procedure?

  4. Each sample retains at least one observed view, whereas training the counterfactual imputation networks requires at least two. At high missing rates, particularly on two-view datasets, how many samples remain available to train these networks? How does the method handle cases with too few co-observed samples?

  5. Table 4 specifies three communication rounds for CFIMC. What communication budgets and stopping criteria are used for the federated baselines? Are all methods evaluated under a common budget, or is each trained to convergence?

9 comments
AuthorsOpenReview ↗

W1 / Q1. FIMCFG, adaptation of HFMVC and FCUIF, imputation for complete-view baselines, centralized references

FIMCFG. We added FIMCFG (ICML 2025) to all datasets. ACC (%), pooled over 5 seeds × 5 missing rates as in Table 1:

DatasetMNAR: FIMCFGMNAR: CFIMCMNAR_latent: FIMCFGMNAR_latent: CFIMC
MSRC-v154.91±7.1578.09±4.0754.84±4.9378.05±5.31
Caltech-5V71.96±5.9082.96±1.8370.31±6.3982.64±2.35
Scene-1537.67±3.8041.56±0.7735.19±5.7641.34±1.03
LandUse-2125.21±2.7426.82±0.9224.02±3.5926.54±0.87
3V-Fashion75.10±9.5874.08±2.7372.39±9.8774.77±5.88
100leaves69.96±12.2272.18±11.7269.18±12.3572.56±12.25
DryBean60.42±8.2685.78±0.1661.55±8.5585.72±0.20
Hdigit76.43±12.1992.72±2.9476.44±15.7792.75±2.79

HFMVC. HFMVC allows several peer clients per view, whose models are aggregated. In our vertical setting, each view is held by one client, the partition used for every method, so the number of peer clients is one: the peer aggregation (its Eq. 2) becomes the identity, while the friend-client contrastive loss, the model-contrastive loss and the heterogeneity-aware choice between its IID and non-IID forms run unchanged on the mean-filled views.

FCUIF. The samples are aligned in our setting, so the alignment module is switched off (alignment rate one, identity alignment matrices), the configuration for fully aligned data. We used the official code with three changes needed for a fair benchmark: it accepts arbitrary masks instead of a complete-first layout; its switch that adopts or discards imputations compared clustering accuracies computed with the ground-truth labels, which we replaced by an unsupervised criterion, the mean silhouette of the view embeddings; and when too few complete samples exist, the global centers are computed on all samples, and the imputer is trained on partially co-observed samples.

Imputation for complete-view baselines. Each client fills the missing rows of its own view with the mean of its visible rows; this replaces the vague footnote. We chose client-local mean imputation because it is the only imputation a client can compute without other clients' data. Any cross-view imputation, by neighbors or regression, would give these baselines access that their vertical protocols do not allow and that FedIMVC methods do not get. It is also the usual way to run complete-view methods on incomplete data (e.g., ICMVC, AAAI 2024).

Centralized IMVC methods (new Appendix H.6, separate table). ICMVC (AAAI 2024), DCG (AAAI 2025) and TreeEIC (TPAMI 2026) use the masks of Table 1, see only observed entries (hidden entries are zeroed and masked) and use their published hyperparameters. They observe all views of a sample at one site, so the table measures the price of federation rather than a comparison under equal conditions. ACC (%), mean ± std over 5 seeds × 5 missing rates:

DatasetMechanismICMVC (AAAI 2024)DCG (AAAI 2025)TreeEIC (TPAMI 2026)CFIMC (federated)
MSRC-v1MNAR59.76±8.5449.57±7.1039.57±5.8478.09±4.07
MSRC-v1MNAR_latent60.10±12.1656.10±10.3944.00±7.9978.05±5.31
Caltech-5VMNAR62.11±6.7870.19±7.2268.19±9.7282.96±1.83
Caltech-5VMNAR_latent60.62±5.4672.54±7.2873.19±12.9382.64±2.35
Scene-15MNAR24.92±1.6938.60±2.3334.54±3.9141.56±0.77
Scene-15MNAR_latent24.26±1.4538.05±1.5334.36±5.4141.34±1.03
LandUse-21MNAR14.39±3.8423.92±1.2122.13±2.3826.82±0.92
LandUse-21MNAR_latent16.14±2.3524.42±1.6520.07±1.7026.54±0.87
3V-FashionMNAR86.15±2.7093.46±2.8666.36±9.1074.08±2.73
3V-FashionMNAR_latent84.45±4.9389.34±5.9967.81±6.6374.77±5.88
100leavesMNAR1.00±0.0062.24±7.4663.79±12.2572.18±11.72
100leavesMNAR_latent1.00±0.0060.80±9.1662.97±12.3772.56±12.25
DryBeanMNAR57.03±6.0664.15±7.2560.55±15.4585.78±0.16
DryBeanMNAR_latent49.89±3.9363.42±7.6551.02±15.2085.72±0.20
HdigitMNAR78.25±19.4383.16±17.0261.02±19.8692.72±2.94
HdigitMNAR_latent81.50±18.5883.73±15.8476.82±18.1292.75±2.79
AuthorsOpenReview ↗

W2 / Q2. Generation of MNAR_latent, definition of η, and the two-view case

The full generator is now given in Appendix H.2. All mechanisms share one two-stage routine:

  1. Which samples are incomplete. Each sample gets a score sis_i, and exactly nη=round(ηN)n_\eta=\mathrm{round}(\eta N) samples are drawn without replacement with probabilities proportional to exp⁡(si)\exp(s_i) (Gumbel-top-k).
  2. Which views are missing. An incomplete sample keeps view mm independently with probability σ(ℓim)\sigma(\ell_{im}). If it would keep all views, one view is dropped with probability ∝1−σ(ℓim)\propto 1-\sigma(\ell_{im}); if it would keep none, one view is kept with probability ∝σ(ℓim)\propto\sigma(\ell_{im}).
  • MNAR: si=γayis_i=\gamma a_{y_i} and ℓim=γbyim\ell_{im}=\gamma b_{y_i m}, with ak,bkm∼N(0,1)a_k,b_{km}\sim\mathcal N(0,1) drawn once per seed and γ=2\gamma=2. Visibility depends on the cluster only, so (P) holds.
  • MNAR_latent (no labels): each view's complete features are standardized column-wise and divided by Dm\sqrt{D_m}, so that every view has the same total weight; the views are concatenated, and the standardized scores on the three leading principal components form a continuous latent factor ui∈R3u_i\in\mathbb R^3. Then si=γ ui⊤a/∥a∥s_i=\gamma\,u_i^\top a/\|a\| and ℓim=γ ui⊤dm\ell_{im}=\gamma\,u_i^\top d_m, with a∼N(0,I3)a\sim\mathcal N(0,I_3) and unit vectors dmd_m drawn once per seed. As uiu_i varies continuously within clusters, so does visibility, and (P) fails. Because uiu_i also loads on view-specific variation, MNAR_latent additionally contains a mild value-dependent component, a further departure from our assumptions, which we now state.

How far MNAR_latent departs from (P) (Table 10). For each view we fit a logistic regression of R(m)R^{(m)} on the one-hot cluster, and one on the cluster and uiu_i.

DatasetI(Y;R) bits: MNAR / MNAR_latentPattern-only ACC %: MNAR / MNAR_latentMajority %ΔR² of R on (Y,u) vs Y: MNAR / MNAR_latent
MSRC-v11.139 / 0.68446.7 / 37.114.30.018 / 0.229
Caltech-5V1.166 / 0.51547.6 / 34.414.30.002 / 0.196
Scene-150.981 / 0.35825.4 / 17.19.10.000 / 0.170
LandUse-211.056 / 0.16020.0 / 9.14.80.001 / 0.210
3V-Fashion1.043 / 0.52839.2 / 22.810.00.000 / 0.133
100leaves1.109 / 0.9455.0 / 4.11.00.006 / 0.033
DryBean0.766 / 0.30742.7 / 34.026.10.000 / 0.134
Hdigit0.708 / 0.23225.2 / 20.610.00.000 / 0.259

η and the two-view case. η is the fraction of incomplete samples (as in FCUIF, and the η=(n−m)/n\eta=(n-m)/n of ICMVC and COMPLETER), not the fraction of missing blocks. The at-least-one-view constraint is enforced in stage 2, when choosing which views to drop, never by discarding or redrawing samples, so the nominal η is attained exactly for every M≥2M\ge2. On two-view datasets every incomplete sample misses exactly one view, so the missing-block fraction is exactly η/2 (45% at η = 0.9), and the co-observed samples are exactly the complete ones. Realized fractions (Table 9):

Missing entries (%), MNAR / MNAR_latent

Dataset (M)η=0.1η=0.3η=0.5η=0.7η=0.9
MSRC-v1 (4)5.6 / 5.115.8 / 15.224.9 / 24.734.1 / 35.642.3 / 45.2
Caltech-5V (5)4.5 / 5.414.0 / 15.924.6 / 26.034.9 / 36.045.6 / 45.4
Scene-15 (3)4.4 / 5.613.7 / 16.224.0 / 26.434.7 / 36.345.4 / 45.8
LandUse-21 (3)4.5 / 5.114.0 / 15.223.8 / 25.133.7 / 35.243.5 / 45.2
3V-Fashion (3)4.6 / 5.014.5 / 15.324.4 / 25.234.4 / 34.943.8 / 44.8
100leaves (3)5.2 / 6.015.4 / 17.325.9 / 27.736.2 / 37.246.3 / 46.3
DryBean (2)5.0 / 5.015.0 / 15.025.0 / 25.035.0 / 35.045.0 / 45.0
Hdigit (2)5.0 / 5.015.0 / 15.025.0 / 25.035.0 / 35.045.0 / 45.0

Realized fraction of incomplete samples minus nominal η: max |dev| = 0.0001 over all 240 masks.

AuthorsOpenReview ↗

W3. The "w/o CF imputation" ablation does not isolate the counterfactual formulation

We agree: replacing Module II by cluster-level imputation changes both the principle and the granularity. Table 2 (Appendix H.7) now keeps Modules I and III fixed and replaces Module II by sample-level imputers with the same cross-fitting folds: the view mean; cross-view kNN (as in DSIMVC and ICMVC); a direct cross-view regressor with the width, epochs and optimizer of Module II (COMPLETER/DCP style); and our own networks trained by plain squared error, without transport weights, InfoNCE, nHSIC and view dropout, which separates the training principle from the architecture. With the hidden views we also report each imputer's oracle R2R^2 on the missing entries and its optimism, i.e., how much higher its transport-weighted cross-fitted R2R^2 on visible entries is. Averages over four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean") and five seeds (η=0.5\eta=0.5); parentheses: paired change vs. full CFIMC.

Module II replaced byGranularityMNAR ACCMNAR NMIMNAR_latent ACCMNAR_latent NMIHidden-entry R²: MNAR / MNAR_latentVisible − hidden R²: MNAR / MNAR_latent
No imputation–66.58 (−1.88)61.73 (−1.06)67.08 (−1.88)60.03 (−2.28)– / –– / –
Cluster-level, FCUIFcluster65.93 (−2.54)61.58 (−1.21)66.71 (−2.25)59.79 (−2.52)0.224 / 0.4070.618 / 0.421
View meanconstant66.47 (−1.99)61.70 (−1.09)67.03 (−1.94)59.92 (−2.39)-0.240 / -0.286-0.185 / 0.081
Cross-view kNNsample66.49 (−1.97)61.43 (−1.36)70.52 (+1.56)64.13 (+1.82)-0.313 / 0.3251.231 / 0.604
Direct cross-view regressionsample68.40 (−0.07)62.63 (−0.16)69.86 (+0.90)62.53 (+0.21)-0.296 / 0.2621.177 / 0.635
Same networks, plain MSEsample68.46 (−0.00)62.97 (+0.18)69.20 (+0.23)62.10 (−0.21)-0.132 / 0.3811.082 / 0.581
Module II (ours)sample68.4762.7968.9762.31-0.233 / 0.3321.189 / 0.634
Reviewer_VTUYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
1(−1)
Read review
Paper summary

This paper investigates federated incomplete multi-view clustering under informative view missingness and proposes CFIMC, a causal framework that explicitly models the missingness mechanism. The method incorporates missingness-aware cluster assignment, sample-level counterfactual imputation, and reliability-weighted prototype learning. The authors provide theoretical analyses and conduct experiments under different missingness mechanisms.

Critical strengths
  1. The paper addresses an important problem that has received relatively limited attention in federated multi-view clustering. Treating missingness patterns as informative signals provides a useful perspective beyond conventional random missingness assumptions.
  2. The proposed framework is coherent, with the missingness model, counterfactual imputation, and reliability estimation addressing different aspects of the problem.
  3. The theoretical analysis is extensive, and the experiments cover multiple datasets and missingness settings, providing substantial support for the proposed approach.
Critical weaknesses
  1. The causal formulation relies on assumptions about the missingness mechanism and data-generating process. The implications of violating these assumptions are not sufficiently discussed.
  2. The framework contains several interacting components, and the specific advantage of causal modeling over simpler missingness-aware strategies could be more clearly established.
  3. The experiments mainly consider constructed missingness mechanisms. The applicability to real federated settings with unknown or changing missingness patterns remains uncertain.
Questions for authors
  1. How sensitive is CFIMC to misspecification of the assumed causal structure or missingness mechanism?
  2. What additional benefits does the causal formulation provide over directly incorporating missingness indicators into clustering?
  3. How would the method handle different missingness mechanisms across clients?
5 comments
AuthorsOpenReview ↗

W1 / Q1. Sensitivity to a misspecified causal structure or missingness mechanism

Thank you for the positive assessment and for the questions on assumptions and mechanisms, which led to the new Section 4.6 and Appendix C.3.

What the algorithm assumes. The algorithm never requires Condition (P); only the results that analyze cluster-level propensities use it, and our MNAR_latent experiments already violate it (visibility varies within clusters; the departure is quantified in Table 10). The new Appendix C.3 (Table 6) goes through every departure and states what still holds:

  • (P) fails. Cluster-level weights still restore each client's cluster proportions, because P(Y=k∣R(m)=1)=πm(k)P(Y=k)/πˉmP(Y=k\mid R^{(m)}=1)=\pi_m(k)P(Y=k)/\bar\pi_m is Bayes' rule and needs no assumption. What remains is selection within clusters, which the doubly robust term reduces by the factor ρˉm\bar\rho_m (new Proposition D.16). Module II is unaffected, since Theorem 3.2 does not use (P).
  • Missingness driven by an always-visible view (MAR). The cluster-level table is misspecified, but for every other view X(m)⊥X(a)∣CX^{(m)}\perp X^{(a)}\mid C, so the invariance behind Module II still holds.
  • Self-masking (a view's own value causes its absence) breaks latent ignorability, and no method that uses only the observed data can identify it (Mohan & Pearl, 2021). The safeguards are the bounded evidence and the mechanism-free bias bound of Proposition D.16(ii).
  • Weak or no information (MCAR, small strength): the table becomes flat, the evidence and weights become inert (Corollaries D.6 and D.21), and CFIMC reduces to its backbone with counterfactual imputation.

In every case, the evidence can overturn a feature-based assignment only when the feature margin is below 2λRσ2Mlog⁡1−π‾π‾2\lambda_R\sigma^2M\log\frac{1-\underline\pi}{\underline\pi} (Corollary D.6), so a wrong mechanism cannot collapse clusters onto missing patterns.

New experiments (Section 4.6, Table 3; per dataset in Appendix H.8). All settings use η=0.5\eta=0.5, so only the mechanism changes, and CFIMC is never told which mechanism generated the mask. ACC (%), averaged over four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean") and five seeds, and I(Y;R) and ΔR² describe the simulated masks:

MechanismI(Y;R) bitsΔR² (P)CFIMCCFIMC w/o mech.FCUIFEFIMVCFIMCFG
MCAR0.1030.00371.4771.5367.4962.0453.73
MAR0.2900.08171.3671.0360.2159.4152.40
MNAR g=0.50.2800.00471.1670.6668.0660.0559.67
MNAR g=10.5990.00467.9967.4860.0462.1762.01
MNAR g=20.9990.00570.2967.1561.2862.3461.64
MNAR g=41.3460.00466.1662.7558.7361.9465.38
Self-masking0.1510.04768.4767.4566.8358.5855.01
Hetero clients0.6260.04767.4766.5263.0560.1056.70
Drift0.4550.00571.1169.8768.9361.1161.08

CFIMC is the best method in 8 of the 9 settings. Its margin over CFIMC w/o mechanism grows with I(Y;R) and vanishes under MCAR, as Corollary D.21 predicts. Under self-masking, which lies outside the identifiable class, it loses 1.82 points relative to MNAR g=2 but remains above / within 1.64 points of the best baseline.

AuthorsOpenReview ↗

W2 / Q2. What the causal formulation adds over feeding missingness indicators to the clustering

Appending the indicators λ ri\sqrt\lambda\,\mathbf r_i to the features is the natural simpler strategy, and we now analyze and test it.

Theory (new Proposition C.6). (i) Indicators change the preference between two clusters with visibilities a,ba,b of a view by λ(a−b)(a+b−2r)\lambda(a-b)(a+b-2r), at most λ\lambda per view. (ii) The likelihood of the pattern, which the MAP rule requires (Proposition D.5), contributes −rlog⁡ab−(1−r)log⁡1−a1−b-r\log\frac ab-(1-r)\log\frac{1-a}{1-b}, and to first order the indicator term equals 2λa(1−a)2\lambda a(1-a) times it. No single λ\lambda is right: indicators under-weight precisely the views that some clusters almost never show, which carry the strongest evidence, and λ\lambda cannot be tuned without labels. (iii) More importantly, indicators act only on the assignment step. They leave the two other consequences of informative missingness untouched: clients train on a selected population (Theorem D.2), and the imputer and the prototypes inherit selection bias (Theorem D.14). Correcting those requires the weights 1/π1/\pi and (1−π)/π(1-\pi)/\pi and the doubly robust correction, that is, a model of why views are missing, not only of which views are missing.

Experiment (Table 2, Appendix H.7). η=0.5\eta=0.5, averages over four datasets ("MSRC-v1", "Caltech-5V", "Hdigit", "DryBean") and five seeds; parentheses give the paired change relative to the full CFIMC. For the indicators we report λ=3\lambda=3, the scale of Λ\Lambda, and the best λ∈{1,3,10}\lambda\in\{1,3,10\} selected with labels, which favors them.

VariantMNAR ACCMNAR NMIMNAR_latent ACCMNAR_latent NMI
Backbone, pattern ignored65.96 (−2.50)62.86 (+0.07)62.00 (−6.96)57.00 (−5.31)
+ missing indicators, λ=365.72 (−2.74)62.51 (−0.28)58.16 (−10.80)53.33 (−8.98)
+ missing indicators, best λ∈{1,3,10}67.37 (−1.10)64.43 (+1.64)61.95 (−7.02)56.23 (−6.08)
+ evidence Λ (ours)62.91 (−5.55)60.85 (−1.94)58.66 (−10.31)53.85 (−8.46)
CFIMC, indicators instead of Λ67.98 (−0.49)63.04 (+0.25)68.72 (−0.24)62.11 (−0.20)
CFIMC w/o mechanism67.15 (−1.32)61.04 (−1.75)69.96 (+1.00)64.01 (+1.70)
CFIMC68.4762.7968.9762.31
AuthorsOpenReview ↗

W3 / Q3. Unknown, changing or client-specific mechanisms in real federated settings

  • Unknown mechanisms. CFIMC never receives the generator. It re-estimates every client's visibility table from the mask in every round, and the mask is already shared in every vertical protocol after entity resolution.
  • Different mechanisms across clients. Each client (view) has its own row of the table and its own weights, so nothing assumes a shared mechanism. A client whose view is missing at random gets a flat row, so its evidence and weights are inert (Corollary D.6 applies view by view), while informative clients contribute. The "Hetero clients" setting above assigns MCAR, MNAR, MNAR_latent and self-masking to the views in turn.
  • Changing mechanisms. In "Drift", 30% of the samples follow reversed class preferences, as if the collection policy had changed. The table then estimates the pooled visibility, and the evidence fades gracefully toward the backbone (Lemma D.4).
  • Natural missingness. We agree that all our mechanisms are simulated. To our knowledge, public multi-view clustering benchmarks with naturally informative missing views and ground-truth labels are not available, which is why all FedIMVC and IMVC baselines simulate masks as well; simulation is also what makes the oracle audits of Appendix H.9 possible. We added this as limitation (3) in the conclusion, with clinical cohorts in which tests are ordered by physicians as the natural testbed, and our code accepts a user-provided mask for such data.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers, AC, and SAC

We thank the three reviewers for their careful and constructive reviews. We appreciate that they found the problem important and underexplored (VTUY, 31UQ, NU6h), the framework coherent, with a clear correspondence between problems, modules, and theory (VTUY, 31UQ, NU6h), and the theoretical analysis extensive (VTUY, 31UQ).

We have uploaded a revised PDF in which all changes are highlighted in blue. The main changes are summarized below; point-by-point replies follow under each review.

Theory

  • New Proposition D.16 (the implemented prototype estimator). It characterizes the estimator we actually run, with the estimated reliability ρ^m\hat\rho_m, the shrinkage, and estimated, possibly wrong, memberships, without Condition (P) and for any missingness mechanism: an exact decomposition of its limit, a mechanism-free bound on the selection terms, and a bound for membership errors. It replaces the over-broad sentence after Theorem 3.4 by a precise statement: an accurate imputer removes the share ρˉm\bar\rho_m of the within-cluster selection bias, and all of it only at ρ=1\rho=1 (31UQ-W4, NU6h-Q1). Remark D.17 states in what sense the prototypes are doubly robust.
  • New Appendix C.3 (Table 6): for every departure from our assumptions (Condition (P) failing, MAR, self-masking, jointly revealed views, client-specific mechanisms, drift, MCAR), which results lose their exactness and what each component still guarantees (VTUY-W1/Q1/Q3).
  • New Proposition C.6: missing-indicator features versus our evidence term Λ\Lambda (VTUY-W2/Q2).

New experiments

  1. FIMCFG (ICML 2025) added to all real-data and synthetic tables (31UQ-W1).
  2. Three recent centralized IMVC methods, ICMVC (AAAI 2024), DCG (AAAI 2025) and TreeEIC (TPAMI 2026), under identical masks and with access only to observed entries, in a separate table (Appendix H.6) (31UQ-W1).
  3. Module II isolated from imputation granularity: sample-level kNN and direct-regression imputers of the same capacity, and our own networks trained by plain squared error, with Modules I and III unchanged, plus the oracle accuracy of every imputer on the hidden entries (Table 2, Appendix H.7) (31UQ-W3, NU6h-W1/Q2).
  4. Missing-indicator features versus the evidence Λ\Lambda, and ρ^m\hat\rho_m versus ρ≡0\rho\equiv0 and ρ≡1\rho\equiv1 (AIPW) (Table 2) (VTUY-W2/Q2, NU6h-Q1).
  5. Mechanisms outside our model: MCAR, MAR, an MNAR strength sweep, self-masking, client-specific mechanisms and a drifting mechanism (new Section 4.6, Table 3) (VTUY-W1/W3/Q1/Q3).
  6. Direct verification of the three mechanisms against the hidden views: information in the patterns and the assignments the evidence changes, client-side selection before and after weighting, and the error of every prototype estimator against oracle prototypes (Appendix H.9) (NU6h-W2).
  7. Protocol statistics: realized missing fractions and the departure from Condition (P) (Appendix H.2), co-observed sample counts at high missing rates (Appendix H.10), communication budgets of all federated methods and an eight-round run of CFIMC (Appendices H.3 and H.11), and the MCAR results (Appendix H.5) (31UQ-W2/Q2–Q5).

Clarifications and corrections

  • Complete specification of the mask generator, including the label-free MNAR_latent and the definition of the missing rate η\eta (Appendix H.2).
  • Configuration of the baselines: client-local mean imputation for complete-view baselines, HFMVC with one client per view, the alignment module of FCUIF disabled, and its label-based adoption switch replaced by an unsupervised criterion (Appendix H.3).
  • TpreT_{\rm pre} and TlocT_{\rm loc} now appear in Algorithm 1, the caption of the synthetic-data table is corrected, and MSRC-v1 is listed with its four views (24, 512, 256, 254).

We hope the revision addresses the reviewers' concerns, and we would be glad to clarify anything further during the discussion.

AuthorsOpenReview ↗

General comment to all reviewers, AC, and SAC

Dear Reviewers, Area Chair, and Senior Area Chair,

As the discussion period comes to a close, we sincerely thank everyone involved in the review of our submission.

We are grateful to Reviewers VTUY, 31UQ, and NU6h for their careful and constructive reviews, which led to substantive improvements: the new Proposition D.16, the analysis of departures from our assumptions, additional baselines, the isolated ablation of Module II, experiments under other missingness mechanisms, and direct audits of the three mechanisms. We especially thank Reviewer 31UQ for the thorough follow-up and the decision to raise the rating, and Reviewer NU6h for confirming that the concerns were addressed.

We also thank the Area Chair for guiding the discussion, and the Senior Area Chair and Program Chairs for organizing the review process.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.