acceptodds
Under review as a conference paper at ICLR 2027

Structure-Aware Conditional Drift Generation for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to discover latent structures from multi-view data with partially unavailable observations. However, existing methods often recover missing information through feature estimation, cross-view averaging, or neighborhood transfer, which may overlook the target-view distribution and weaken sample correspondence. Although recent approaches improve missing-view recovery by exploiting cross-view consistency or local relationships, they generally provide limited joint modeling of local geometry and cluster-level semantics, while the reliability of recovered representations is rarely considered during subsequent fusion. To tackle these issues, we propose Structure-Aware Conditional Drift Generation (SACDG) for IMVC. Specifically, SACDG takes shared semantics from observed views as a reliable cross-view reference and learns a structure-guided drift toward the target-view distribution, thereby preserving sample correspondence while adapting to target-specific geometry and cluster semantics. Additionally, pseudo-missing self-supervision estimates recovery reliability to reduce the influence of inaccurate completions during multi-view fusion. Finally, structural alignment and prototype-driven clustering provide consistent feedback for representation recovery and clustering. Extensive experiments on seven benchmark datasets under multiple missing rates demonstrate the effectiveness and robustness of SACDG compared with state-of-the-art IMVC methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-view Clustering, Conditional Drift Generation, Graph Structure Alignment, Prototype-driven Clustering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YujwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses Incomplete Multi-View Clustering (IMVC) and proposes Structure-Aware Conditional Drift Generation (SACDG). Using the cross-view semantic anchor constructed from observed views as its basis, the method combines local graph context, prototype semantics, and random perturbations. A conditional drift generator learns target-view-specific residuals to recover the local structure and clustering semantics of missing views while preserving sample correspondences. The model then estimates recovery reliability through pseudo-missing self-supervision and uses confidence and reliability-aware weights for missing-representation calibration and multi-view fusion, while jointly optimizing graph-structure alignment and prototype-driven clustering. The experimental evaluation is comprehensive, and the results provide convincing evidence of the effectiveness and stability of the proposed method.

Critical strengths
  1. The core motivation is well aligned with the method design and clustering prototypes provide constraints at different levels, giving the recovery process a clear methodological logic.
  2. Reliability estimation is tightly connected to missing-representation recovery and multi-view fusion. The paper constructs a reliability target from pseudo-missing samples to supervise confidence. Thus, recovery reliability directly affects later fusion, making the overall design coherent.
Critical weaknesses

1.The core concept of Conditional Drift could be defined more clearly. 2. The authors should unify the relevant notation and definitions to avoid ambiguity.

Questions for authors
  1. In Eq. (7), the cross-view anchor aiva_i^v is obtained by simply averaging all observed views. Could the authors explain the rationale for this choice? For example, is aiva_i^v intended as a relatively neutral semantic starting point, with reliability differences across views handled later during fusion through ωiv\omega_i^v?
  2. How is the random perturbation in Eq. (9) used during training and inference? Please clarify its sampling frequency (e.g., whether it is resampled at every forward pass) and whether clustering uses random sampling, fixed sampling, or a deterministic setting, to make the conditional drift generation process easier to reproduce.
  3. The construction of the pseudo-missing set SS is described only briefly. Please specify the pseudo-missing ratio, whether different views are sampled in a balanced manner, and whether SS is resampled at each epoch or iteration.
  4. Section 3.6 states that LrecL_rec is first used for warm-up and then replaced by the joint objective LL. Could the authors provide the warm-up stopping condition or number of epochs, and explain how prototype PP and soft assignment QQ are initialized when joint optimization begins? This would help readers accurately understand the joint optimization procedure.
6 comments
AuthorsOpenReview ↗

W1. Definition of Conditional Drift

W1: We appreciate the reviewer’s careful assessment of the conceptual formulation and the opportunity to clarify what conditional drift means in ATDA. Conditional drift is a target-directed displacement in the shared latent space, learned from a sample-specific cross-view anchor under complementary structural, prototype-semantic, and stochastic conditions. It is not a temporal drift process and does not refer to the forward or reverse process of a diffusion model. For sample ii and target view vv, ATDA first forms an anchor aiva_i^v from the observed source-view representations, then predicts a residual displacement Δiv\Delta_i^v conditioned on the anchor and the three types of information: $ \Delta_i^v=f_{\theta_v}(a_i^v,s_i^v,p_i,\epsilon_i^v),\qquad \widetilde h_i^v=a_i^v+\Delta_i^v, $ where sivs_i^v denotes the graph context, pi=∑k=1KqikPkp_i=\sum_{k=1}^{K}q_{ik}P_k is the prototype-semantic cue, and ϵiv\epsilon_i^v is a stochastic latent perturbation. Thus, “drift” is the learned target-directed displacement from shared semantics toward the missing view, not unconstrained prediction from scratch.

The anchor and the conditions have distinct roles. Let Ωi−v\Omega_i^{-v} denote the set of views observed for sample ii, excluding the target view vv. The anchor is their arithmetic mean, aiv=∣Ωi−v∣−1∑u∈Ωi−vhiua_i^v=|\Omega_i^{-v}|^{-1}\sum_{u\in\Omega_i^{-v}}h_i^u. It preserves instance correspondence and provides a common semantic starting point. The graph context sivs_i^v aggregates the K-nearest-neighbor representations of the same observed source views, thereby describing the local manifold around sample ii rather than relying only on its own feature vector. In the implementation, the neighborhood size is kg=12k_g=12. The prototype cue pip_i contributes cluster-level information through the current soft assignment qikq_{ik} and prototype PkP_k. The perturbation ϵiv∼N(0,Id)\epsilon_i^v\sim\mathcal N(0,I_d) exposes the generator to stochastic variations around the anchor during training. These signals are complementary: the anchor identifies the sample, the graph context describes its local geometry, the prototype cue supplies cluster semantics, and the perturbation supports conditional generation under uncertainty.

The target-view index is also reflected in the generator parameters. ATDA instantiates a separate generator for each target view, denoted by fθvf_{\theta_v}, rather than applying one shared parameter set to every missing view. The same sample may therefore use common observed evidence to recover multiple missing views, while each generator learns a different target-specific displacement. The architecture processes the anchor, graph context, and prototype cue as condition tokens through multi-head cross-attention. A gated residual network then predicts Δiv\Delta_i^v, and a confidence head estimates the reliability of the prediction. Thus, “conditional” refers to the explicit dependence of the displacement on sample correspondence, local structure, cluster semantics, and stochastic input; “drift” refers to the learned target-directed residual added to the anchor.

When several views are missing, each target is recovered from the available observed-source set and its own target-specific generator; a missing view is never treated as observed evidence for recovering another missing view. If only one source view is available, the anchor reduces to that view's latent code, while the graph and prototype conditions still provide complementary guidance. This covers different missingness patterns.

The residual formulation also explains why the anchor is retained in the output. Directly predicting h~iv\widetilde h_i^v would ask the generator to reconstruct both shared sample semantics and target-specific information simultaneously. Predicting Δiv\Delta_i^v instead separates these roles: the anchor carries the common component, while the generator focuses on the target-view-specific adjustment. This reduces the burden on the generator and provides an interpretable reference for the subsequent confidence calibration and reliability-aware fusion. The generated representation is therefore not simply an average of observed views; the average initializes the recovery, while graph and prototype conditions guide its adaptation toward the target distribution.

For clarity and consistency, we use aiva_i^v for the cross-view anchor, sivs_i^v for the graph context, PkP_k for the kk-th prototype, qikq_{ik} for its soft assignment, pip_i for the resulting prototype cue, ϵiv\epsilon_i^v for Gaussian perturbation, Δiv\Delta_i^v for the predicted displacement, and h~iv\widetilde h_i^v for the generated target-view representation. The target-specific parameters are denoted by θv\theta_v. With these definitions, conditional drift is a structured, anchor-relative generation mechanism that moves shared semantics toward the target-view distribution while preserving instance correspondence and incorporating local geometry and cluster-level guidance.

AuthorsOpenReview ↗

Q1. Rationale for the Cross-View Anchor

Q1: We appreciate the reviewer’s question about the rationale for the arithmetic mean in Eq. (7). The cross-view anchor is designed as a neutral, sample-level semantic reference, rather than as a reliability-weighted estimate of the missing target representation. For sample ii and target view vv, let Ωi−v\Omega_i^{-v} be the set of observed source views excluding vv. The anchor is $ a_i^v=\frac{1}{|\Omega_i^{-v}|}\sum_{u\in\Omega_i^{-v}}h_i^u. $ All view-specific encoders map their inputs to the same latent dimension dd, so the observed representations can be aggregated in a common space. Arithmetic averaging gives each available source view an equal contribution and avoids introducing an additional view-selection network or a learned weighting rule at the anchor-construction stage.

This choice is motivated by the role assigned to the anchor in conditional drift. It preserves the identity and shared semantics of the same sample across views, while leaving target-specific adaptation to the conditional generator. It is not intended to claim that all views are equally informative, nor to assume that their feature distributions are identical. Instead, the mean provides a stable semantic reference that does not privilege one source view before the target-specific evidence has been evaluated. The generator receives this reference together with the graph context sivs_i^v, prototype cue pi=∑k=1KqikPkp_i=\sum_{k=1}^{K}q_{ik}P_k, and stochastic perturbation ϵiv\epsilon_i^v. It learns a target-directed displacement from the anchor, so differences between source views can be reflected in the structural and semantic conditions and in the learned mapping.

The anchor is also target-aware in its construction. When generating view vv, the target view is excluded from Ωi−v\Omega_i^{-v}, and only the other observed views contribute. Consequently, no unavailable target feature is used to form the anchor. If one source view is observed, the mean reduces to that source view's latent representation. If multiple source views are observed, the mean combines them into a common starting point. If the target is missing and the sample retains at least one other observed view, the construction remains well-defined. This property allows the same formulation to cover different missing-view patterns without adding a separate anchor rule for each pattern.

Reliability differences are handled after the target-view prediction, rather than being embedded in the anchor. The confidence score civc_i^v calibrates the generated representation in Eq. (10), controlling how strongly the prediction is retained relative to the anchor. The fusion weight ωiv\omega_i^v then adjusts the contribution of each completed view in Eqs. (16) and (17). In addition, Eq. (20) uses reliability information to reduce the influence of uncertain completed representations during alignment. These stages allow ATDA to distinguish between two questions: what shared semantic reference is available for a sample, and how reliable is the representation generated for a particular target view? Keeping these functions separate makes the anchor simple and interpretable while allowing the downstream modules to account for sample- and view-dependent uncertainty.

A reliability-weighted anchor would combine semantic aggregation and confidence estimation before generation. ATDA instead establishes correspondence with the mean and learns confidence from pseudo-missing recovery supervision. Observed sample-view pairs provide known targets for training the confidence head; at inference, calibrated confidence and fusion weights determine each generated view's contribution to the consensus. Reliability is therefore applied directly to the generated target and its subsequent use.

The arithmetic mean also provides a consistent reference across samples with different numbers of observed views. Its normalization by ∣Ωi−v∣|\Omega_i^{-v}| prevents the magnitude of the anchor from growing simply because more source views are available. The accompanying source-count and dispersion descriptors provide additional information about observation coverage and disagreement to the generator, without changing the definition of the anchor. This is important because two samples can have the same mean but different numbers or dispersion of observed source representations. The generator can use these descriptors, graph context, and prototype cue to adapt its displacement accordingly.

Therefore, the anchor is intentionally a neutral semantic starting point, not a final estimate of the target view and not a measure of source reliability. The conditional drift generator supplies target-specific adaptation, while civc_i^v, ωiv\omega_i^v, and the reliability-aware alignment objective handle uncertainty in the completed views. This separation preserves a clear division of roles and allows ATDA to combine simple sample-level correspondence with target-specific generation and reliability-aware fusion.

AuthorsOpenReview ↗

Q2. Gaussian Perturbation and Reproducibility

Q2: We appreciate the reviewer’s emphasis on the sampling details needed to reproduce conditional drift generation. In the generator, the perturbation is sampled as a standard Gaussian tensor with the same shape as the anchor on every training forward call: $ \epsilon_i^v\sim\mathcal N(0,I_d),\qquad a_{i,\mathrm{noise}}^v=a_i^v+\sigma\epsilon_i^v, $ where σ=min⁡(softplus⁡(ℓ),0.50)\sigma=\min(\operatorname{softplus}(\ell),0.50) is controlled by a learnable scalar ℓ\ell, initialized to −2.5-2.5. The implementation uses torch.randn_like(anchor) whenever the generator is in training mode. Thus, a fresh perturbation is drawn for each forward call and each target-view generator; it is not fixed for an entire epoch, and no predefined sequence of noise levels is traversed. The random tensor is scaled before being added to the anchor and is used as part of the generator's seed input.

Each target view has an independently instantiated generator and therefore its own learnable noise-scale parameter. The parameter is initialized at −2.5-2.5 in its unconstrained form; applying softplus gives a small positive scale at initialization, and the upper clamp prevents the perturbation magnitude from exceeding 0.500.50. The sampled tensor has the latent dimension dd for every sample, and each coordinate is drawn from the standard normal distribution. Importantly, the noisy anchor is used in the generator's seed projection, whereas the cross-attention condition tokens and the residual output retain access to the unperturbed anchor. The perturbation therefore enriches the query condition without replacing the semantic reference or directly corrupting the reconstruction target.

The perturbation is a stochastic condition for learning a target-directed displacement, not a diffusion variable. ATDA does not define a diffusion time index, forward noising schedule, noise-prediction loss, or iterative reverse-denoising trajectory. The generator uses the perturbed anchor to form its query representation, while the condition tokens include the unperturbed anchor, graph context, and prototype-conditioned representation. The output remains an anchor-relative prediction, h~iv=aiv+Δiv\widetilde h_i^v=a_i^v+\Delta_i^v. In this way, stochastic variation enriches the generator's training conditions without replacing the stable semantic reference or changing the target of reconstruction.

At inference, the model is placed in evaluation mode, and the generator uses torch.zeros_like(anchor) instead of drawing random noise. Therefore, ϵiv=0\epsilon_i^v=0 and the perturbed anchor equals the original anchor. The learned scale σ\sigma does not introduce stochastic variation at inference. Each missing representation is produced directly from the anchor, graph context, and prototype cue, followed by confidence calibration and reliability-aware fusion. There is no repeated noise removal, reverse sampling, or random perturbation at inference. This train-evaluation distinction makes the prediction path deterministic for a fixed input, trained parameter set, graph, and prototype state. The zero-noise setting is applied consistently to every target-view generator, so inference does not depend on a newly drawn latent vector or on the number of sampling iterations.

Clustering also does not use random sampling to obtain the soft assignment matrix during joint optimization. Once the consensus prototypes are initialized, the assignment for a fused representation ziz_i is calculated by normalized similarity followed by softmax: $ q_{ik}=\frac{\exp(\operatorname{sim}(\bar z_i,\bar P_k)/\tau)} {\sum_{r=1}^{K}\exp(\operatorname{sim}(\bar z_i,\bar P_r)/\tau)}, $ where zˉi\bar z_i and Pˉk\bar P_k are normalized representations and prototypes, and τ\tau is the assignment temperature. Thus, QQ is computed deterministically from the current fused representation and prototype parameters; it is not obtained by sampling cluster labels. The prototype initialization uses the configured seed, and the training code propagates that seed to Python, NumPy, and PyTorch while disabling cuDNN benchmarking and enabling deterministic cuDNN behavior.

Gaussian perturbation and pseudo-missing selection are stochastic training operations, whereas soft cluster assignments are deterministic given the current representations and prototypes. K-means initialization uses a seeded generator, while Python, NumPy, and PyTorch receive the configured experiment seed; CUDA cuDNN benchmarking is disabled and deterministic execution is enabled. Thus, ATDA does not randomly sample QQ or run an inference-time denoising chain.

The same setup governs repeated runs at each missing rate, so the stochastic training masks and perturbations follow a controlled random stream. This design provides stochastic regularization during training and a deterministic prediction path at inference. Together with seeded initialization and deterministic soft assignments, it specifies the random and non-random parts of the procedure.

Reviewer_8aHROpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors think existing methods often ignore the reliability of recovered representations during subsequent fusion. To this end, they propose ATDA (their new title) for IMVC. This method uses a cross-view semantic anchor as the starting point for missing-view generation and learns a target-view-specific conditional drift under local graph context and shared prototype semantics. Extensive experiments on seven benchmark datasets demonstrate its effectiveness.

Critical strengths
  1. The authors formulate recovery as learning a target-view-specific drift from a reliable cross-view semantic anchor, with graph context and prototype semantics jointly guiding the shift toward the target-view distribution.
  2. The paper is well organized and easy to read. The experiments are comprehensive and well designed.
Critical weaknesses
  1. How does the model distinguish different target views and learn view-specific distributions? Please clarify whether the generator uses shared or view-specific parameters.
  2. Eq. (9) is similar to conventional conditional residual prediction. Please explain the difference between your drift mechanism and the standard residual generator. In addition, please clarify the role of Gaussian perturbations in distribution modeling.
  3. Missing-views are recovered by the current prototypes, and the prototypes are updated by the recovered views. It seems that errors will be bigger and bigger along with the training. How do you address this issue? I suggest that you provide further analysis or stability experiments.
  4. The original title is similar to the TKDE paper “Structure-Aware Conditional Diffusion Generation for Incomplete Multi-View Clustering.” Please clarify the relationship between these two papers. If your work is an extension, please tell me the contributions and confirm that it satisfies the ICLR’s anonymity policy. If not, please cite the TKDE paper and explain the differences between them.
Questions for authors

See weaknesses

7 comments
AuthorsOpenReview ↗

W1: Target-View-Specific Conditional Drift Generation

W1: We appreciate the reviewer’s careful reading of our work and the opportunity to elaborate on the target-view-specific design of ATDA. The target-view identity is encoded in the parameterization of the conditional drift generator. Appendix B.1 specifies that a generator is instantiated independently for each target view. Thus, the compact notation fθf_{\theta} in Eq. (9) represents a family of target-view-specific generators. With the target index made explicit, the expression is $ \widetilde h_i^v=a_i^v+f_{\theta_v}\left(a_i^v,s_i^v,\sum_{k=1}^{K}q_{ik}P_k,\dot q\right), $ where vv denotes the target view and θv\theta_v denotes the parameters of its generator. ATDA adopts a consistent architectural design across these modules, while retaining independent parameters for different target views.

The conditions provide complementary recovery information, while generator parameters adapt it to the target distribution. The cross-view anchor aiva_i^v aggregates the latent representations of the views observed for sample ii, preserving shared sample semantics and providing a stable reference. The graph context sivs_i^v aggregates KK-nearest-neighbor representations from observed source views and describes local geometry. The prototype cue ∑kqikPk\sum_k q_{ik}P_k summarizes cluster-level semantics according to the current soft assignment. These conditions may use the same observed source set when several target views are missing, because they represent the evidence available for that sample. Sharing available evidence does not imply sharing the target mapping.

Target-specific mappings are learned through separate pseudo-missing supervision. During training, an observed sample--view pair is temporarily masked, the remaining observed views construct the anchor and graph context, and the hidden view's latent representation provides the recovery target. The generator associated with that view is optimized using these target-specific pairs. Consequently, θv\theta_v is shaped by the feature distribution and recovery targets of view vv, rather than by a pooled objective that treats all target views as interchangeable. When a sample has multiple missing views, ATDA invokes the corresponding generators separately. For example, missing views v1v_1 and v2v_2 are recovered by fθv1f_{\theta_{v_1}} and fθv2f_{\theta_{v_2}}, respectively. Even if both predictions use the same observed source information, their learned transformations and supervision targets differ, allowing their displacements to reflect distinct target-view characteristics.

A common latent dimension makes views compatible for comparison and fusion; it does not require identical distributions. ATDA retains view-dependent information in the feature stems, adapters, decoders, and target-view-specific drift generators. The common latent space enables interaction among anchors, graph context, and prototypes, while each target-specific mapping preserves view-dependent adaptation. This is particularly important when different views encode different aspects of the same sample: their semantic starting point can be shared without forcing their recovered features to follow the same target distribution.

The observation set Ωi\Omega_i determines which available views contribute to the anchor and graph context. Every sample retains at least one observed view, so recovery conditions are constructed without accessing unavailable target features. Each missing target is generated by its own view-specific module. The number of missing views changes how many target-specific generators are invoked, not their identity or parameterization.

The generator architecture further distinguishes ATDA from a simple view-conditioned correction. The anchor, graph context, and prototype semantics are projected into latent tokens and interact through multi-head cross-attention. The attended representation passes through a residual MLP, a sigmoid gate controls the predicted drift magnitude, and a separate confidence head estimates recovery reliability. Thus, target-specific parameters operate on structured conditions encoding sample correspondence, local geometry, and cluster semantics. Confidence calibration controls how much generated information is retained relative to the anchor, protecting subsequent fusion from unreliable predictions.

Accordingly, Eq. (9) is a compact expression for target-view-specific generators, not one parameter-shared generator applied to every missing view. Appendix B.1 specifies this implementation, and pseudo-missing training supplies target-specific supervision. The anchor and graph context exploit information shared by the observed part of a sample, while θv\theta_v learns the displacement appropriate for each missing target. This design enables separate recovery of multiple missing views while maintaining cross-view semantic correspondence, local structural continuity, and cluster-level consistency. This separation is also maintained at inference.

AuthorsOpenReview ↗

W2: Conditional Drift Modeling and Gaussian Perturbation

W2: We are grateful for the reviewer’s thoughtful evaluation of the conditional drift mechanism and the opportunity to explain its design principles. The additive form in Eq. (9) preserves the cross-view anchor as a stable reference; ATDA's contribution does not rest on addition alone. It lies in how the target-view displacement is conditioned, generated, calibrated, and integrated into incomplete multi-view clustering. It jointly models sample correspondence, geometry, cluster semantics, and reliability.

A conventional residual MLP learns a correction from its input, but the residual form alone does not specify how to preserve sample identity, follow the local manifold, or remain consistent with cluster structure. ATDA explicitly incorporates these three sources of information. The cross-view anchor aiva_i^v aggregates observed latent representations of the same sample and provides an instance-level semantic reference. The graph context sivs_i^v aggregates KK-nearest-neighbor representations from observed source views and introduces local geometry beyond the anchor. The prototype cue ∑k=1KqikPk\sum_{k=1}^{K}q_{ik}P_k injects cluster semantics according to the current soft assignment. Projected into latent tokens, these conditions interact through multi-head cross-attention and combine complementary evidence.

The attended feature passes through a residual MLP, a sigmoid gate controls displacement magnitude, and a separate confidence head estimates reliability. Each target view has its own generator and pseudo-missing supervision. The anchor-relative formulation predicts a target-directed displacement under graph and prototype guidance, rather than an unconstrained correction. Therefore, the residual parameterization is only the output form; the conditioning and reliability pathways define the actual recovery mechanism.

Gaussian perturbation q˙∼N(0,Id)\dot q\sim\mathcal N(0,I_d) is a stochastic latent condition introduced during training to model uncertainty around the anchor. It exposes the generator to perturbed conditions while retaining the anchor as its semantic reference. It is not a regression target or a diffusion state: ATDA has no diffusion time index, prescribed noise schedule, a noise-prediction objective, or an iterative reverse-sampling procedure. Recovery reliability is learned separately from pseudo-missing reconstruction errors. Gaussian perturbation supports conditional generation; the main contribution is structure- and prototype-conditioned recovery with reliability-aware clustering.

The reliability pathway connects recovery to downstream use. During training, observed sample--view pairs are temporarily hidden to form pseudo-missing targets. Their recovery errors define continuous reliability targets $ \rho_i^v=\exp\left(-|\widetilde h_i^v-h_i^v|_2^2/\kappa_v\right), $ which supervise the confidence head. Confidence controls interpolation between the generated representation and the anchor. During fusion, observed views receive unit unnormalized weight, whereas missing views are weighted by confidence, restraining unreliable generations at both stages. The reliability-weighted fusion objective scales each view's alignment contribution, while topology preservation maintains neighborhood relations in the fused space.

A parameter-matched residual baseline and isolated noise ablation would compare generic generator variants or quantify one stochastic input. These comparisons address a narrower component-level question than ATDA's central IMVC task. Our focus is the integrated task: target-specific recovery, local-geometry preservation, reliability calibration, and robust clustering. A generator-only comparison would not evaluate component interactions, while a noise-only ablation would not measure the complete anchor-to-target and reliability-aware framework. We therefore evaluate ATDA at the level of its intended task and integrated objectives, rather than positioning it as a study of generic residual backbones or noise parameterization. The objective-level ablation directly matches this scope by comparing the full model with variants that remove the completion objective, the graph-fusion-clustering objective, or both.

The existing experiments provide broad evidence for this integrated design. ATDA is compared with eight recent IMVC methods on seven datasets under four primary missing rates, using ACC, NMI, and ARI. The full objective consistently outperforms reduced variants on the reported ablation datasets. The supplement reports additional missing-rate results, parameter sensitivity and convergence on four datasets under five rates, and representation visualizations. These evaluations examine end-to-end behavior under incomplete-view conditions across datasets. ATDA's central advantage is the coordinated use of target-specific conditional drift, graph and prototype guidance, pseudo-missing reliability calibration, and reliability-aware structural fusion.

AuthorsOpenReview ↗

W3: Coordinated Optimization of Prototype Learning and Missing-View Completion

W3: We sincerely appreciate the reviewer’s valuable comments and the opportunity to elaborate on the coordinated learning strategy of ATDA. The interaction between missing-view recovery and prototype learning is an intentional component of our framework: completion informs fusion, fusion updates prototypes, and prototype semantics guide subsequent recovery. Staged optimization, progressive activation, pseudo-missing supervision, confidence calibration, high-confidence self-training, and topology preservation organize this feedback.

ATDA first separates representation warm-up from joint recovery and clustering. In the base configuration, view encoders and decoders are warmed up for 7070 epochs using LrecL_{\mathrm{rec}}, which combines Smooth-L1 feature reconstruction with a structural term comparing pairwise similarities of reconstructed and original features. Only observed entries contribute to this stage. This establishes view-specific representations and local structure before joint recovery and clustering. The subsequent joint stage uses 220220 epochs and optimizes the complete objective in Eq. (24) with AdamW, weight decay 10−510^{-5}, gradient clipping at 55, and cosine annealing. Dataset profiles adjust dimensions and training length while retaining this procedure.

The missing-view branch is introduced progressively rather than receiving full influence from the first joint epoch. Its progress coefficient is $ \alpha_t=\min\left(1,\frac{t+1}{T_{\mathrm{imp}}}\right), $ where TimpT_{\mathrm{imp}} is the imputation warm-up length. Pseudo-missing entries are sampled from currently observed entries at the default rate of 0.250.25. Their known latent representations supervise LpredL_{\mathrm{pred}}, combining Smooth-L1 recovery and cosine-direction consistency, and provide recovery-error-based reliability targets for confidence supervision. The generator is optimized against observed latent targets, and confidence reflects recovery quality.

Prototype-based clustering is delayed by a short warm-up and introduced through a gradual ramp. In addition, the clustering loss does not treat every current assignment as a reliable pseudo-label. Only samples satisfying max⁡kqik≥ξ\max_k q_{ik}\geq\xi contribute to LcluL_{\mathrm{clu}}. This selection limits the influence of uncertain pseudo-labels, while the complete soft-assignment matrix remains available for calculating smooth cluster centers. Thus, uncertain early assignments do not dominate feedback.

The prototype update is an assignment-weighted center of the fused representations: $ P_k=\frac{\sum_iq_{ik}z_i}{\sum_iq_{ik}}. $ Appendix A.5 proves that, for fixed ZZ and QQ with ∑iqik>0\sum_iq_{ik}>0, this update is the unique minimizer of $ J_k(p)=\sum_iq_{ik}|z_i-p|_2^2. $ Each update is therefore the soft center induced by current assignments, rather than an arbitrary movement independent of representation geometry.

Integration further controls feedback. An observed view has unnormalized fusion weight one, whereas a missing view is weighted by its predicted confidence civ∈[0,1]c_i^v\in[0,1]. Confidence calibration also interpolates between the generated code and cross-view anchor, keeping uncertain predictions closer to observed semantics. In addition, LfusL_{\mathrm{fus}} weights view-to-fusion consistency by reliability, and LtopL_{\mathrm{top}} preserves neighborhood similarities between each view-specific graph representation and the fused space. These constraints control generated-view contributions and local geometry.

The paper evaluates optimization under incomplete views. Figure 3 reports convergence on HW, MNIST4, MNIST-Fea, and Synthetic3D. Appendix C.5 and Figure 8 extend the analysis to five missing rates, 0.10.1, 0.30.3, 0.50.5, 0.70.7, and 0.80.8, across all four datasets. The curves show progressive improvement followed by stable regions; increasing missingness makes optimization more challenging but does not fundamentally change the overall convergence pattern. Appendix C.4 also reports parameter sensitivity on the same datasets and rates, showing broad favorable regions rather than dependence on a narrowly selected combination. Objective ablation further assesses the complete method under incompleteness.

Isolated prototype trajectories focus on one internal variable, whereas ATDA addresses end-to-end clustering with interacting completion, graph, and cluster semantics. The staged schedule, confidence-controlled fusion, weighted-center update, topology constraint, and convergence results support this objective. We therefore focus on the integrated stability mechanisms and observed clustering behavior, rather than expanding the scope to a separate study of generic prototype-dynamics diagnostics. The feedback between completion and prototypes is deliberate, structurally constrained at the update level, and empirically evaluated across datasets and missing rates. This coordinated design is evaluated as a complete clustering framework. This is intentional.

Reviewer_CdY9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Motivated by the tendency of missing-view imputation to deviate from sample semantics, distort local structures, and propagate errors during fusion, the paper proposes Structure-Aware Conditional Drift Generation (SACDG). The framework consists of three main components: view representation learning with structure preservation, structure-aware conditional drift completion with reliability estimation, and graph structure alignment with prototype-driven clustering. Its main contribution is to learn imputation residuals from semantic anchors derived from observed views, guided by neighborhood and prototype information, while using confidence to control both the magnitude of the imputation adjustment and the fusion weights. This enables joint optimization of imputation and clustering.

Critical strengths

S1: When the observed views already provide sufficient information, the generator can produce a small correction; when more target-view information is needed, it can learn a larger adjustment. This design preserves a direct path for passing through existing sample information and allows the generator to focus its learning capacity on cross-view differences.

S2: The authors construct a pseudo-missing task by temporarily masking observed sample–view pairs, allowing each masked pair to provide supervision for both imputation prediction and reliability estimation. This design gives confidence estimation an observable learning signal without requiring additional annotations of which imputed representations are reliable.

Critical weaknesses

W1: Equation (9) primarily describes conditional residual prediction, which the authors refer to as “conditional drift.” However, the current manuscript does not sufficiently establish its substantive distinction from a conventional conditional residual network. W2: Even if confidence supervision is implemented correctly, Equation (12) still measures the Euclidean distance to the current target latent representation, whereas fusion requires assessing whether the representation contributes to correct clustering. These two criteria are not fully equivalent. W3: Equation (10) uses confidence during imputation, and Equation (16) uses it again during fusion. However, the neighborhood aggregation in Equation (15) takes a simple average without further weighting neighbors by the reliability of their imputed representations.

Questions for authors

Q1: Compared with a direct MLP or a deterministic residual MLP with matched parameter counts and training budgets, what are the methodological differences and empirical advantages of the proposed approach? How does performance change when the random perturbation is removed? Q2: After masking the target view, are the anchor, neighborhood conditions, and prototype assignments recomputed to prevent target information from entering the generation conditions? How are samples with only one observed view handled?

1 comment

Thank you for the author's detailed response; I will keep my score as is.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

Dear PCs, SACs, ACs, and Reviewers,

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

In response to the reviewers' concerns, we have substantially revised our manuscript, refined the presentation of our proposed method, and conducted additional experiments and analyses. The major updates are summarized below:

We revised the paper title to Anchor-to-Target Drift Adaptation for Incomplete Multi-View Clustering and consistently updated the method name to ATDA, better reflecting the central idea and distinguishing our approach from related existing studies.

We further clarified the formulation and motivation of conditional drift adaptation, elaborating on the complementary roles of cross-view semantic anchors, local graph contexts, prototype guidance, and stochastic perturbations in target-view recovery.

We expanded the experimental evaluation to seven benchmark datasets and eight representative IMVC baselines, covering eight missing rates from 0.1 to 0.8 and reporting ACC, NMI, and ARI.

We conducted additional objective ablation experiments on multiple datasets to investigate the individual and joint contributions of reconstruction, conditional drift completion, graph alignment, and prototype-driven clustering.

We added comprehensive parameter sensitivity and convergence analyses under different missing rates, together with additional representation visualizations to evaluate the robustness of ATDA under increasingly incomplete observations.

We further strengthened the theoretical explanations and implementation details, including the optimization procedure, pseudo-missing self-supervision, confidence estimation, inference strategy, and computational complexity.

All revisions have been incorporated into the updated manuscript, with changes highlighted in blue. Detailed explanations and responses to individual reviewer concerns are provided in our point-by-point rebuttal.

We sincerely hope that these revisions adequately address the reviewers' concerns and clarify the novelty, effectiveness, and robustness of our proposed approach. We would greatly appreciate any further feedback from the reviewers during the discussion period.

Thank you again for your time and consideration.

Sincerely,

The Authors of Submission

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.