- 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
- 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.
- 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
- In Eq. (7), the cross-view anchor is obtained by simply averaging all observed views. Could the authors explain the rationale for this choice? For example, is intended as a relatively neutral semantic starting point, with reliability differences across views handled later during fusion through ?
- 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.
- The construction of the pseudo-missing set is described only briefly. Please specify the pseudo-missing ratio, whether different views are sampled in a balanced manner, and whether is resampled at each epoch or iteration.
- Section 3.6 states that is first used for warm-up and then replaced by the joint objective . Could the authors provide the warm-up stopping condition or number of epochs, and explain how prototype and soft assignment are initialized when joint optimization begins? This would help readers accurately understand the joint optimization procedure.
6 comments
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 and target view , let be the set of observed source views excluding . 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 , 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 , prototype cue , and stochastic perturbation . 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 , the target view is excluded from , 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 calibrates the generated representation in Eq. (10), controlling how strongly the prediction is retained relative to the anchor. The fusion weight 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 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 , , 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.
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 is controlled by a learnable scalar , initialized to . 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 in its unconstrained form; applying softplus gives a small positive scale at initialization, and the upper clamp prevents the perturbation magnitude from exceeding . The sampled tensor has the latent dimension 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, . 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, and the perturbed anchor equals the original anchor. The learned scale 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 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 and are normalized representations and prototypes, and is the assignment temperature. Thus, 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 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.
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 i and target view v, ATDA first forms an anchor aiv from the observed source-view representations, then predicts a residual displacement Δiv 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 siv denotes the graph context, pi=∑k=1KqikPk is the prototype-semantic cue, and ϵiv 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 denote the set of views observed for sample i, excluding the target view v. The anchor is their arithmetic mean, aiv=∣Ωi−v∣−1∑u∈Ωi−vhiu. It preserves instance correspondence and provides a common semantic starting point. The graph context siv aggregates the K-nearest-neighbor representations of the same observed source views, thereby describing the local manifold around sample i rather than relying only on its own feature vector. In the implementation, the neighborhood size is kg=12. The prototype cue pi contributes cluster-level information through the current soft assignment qik and prototype Pk. The perturbation ϵiv∼N(0,Id) 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 fθv, 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, 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 hiv would ask the generator to reconstruct both shared sample semantics and target-specific information simultaneously. Predicting Δiv 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 aiv for the cross-view anchor, siv for the graph context, Pk for the k-th prototype, qik for its soft assignment, pi for the resulting prototype cue, ϵiv for Gaussian perturbation, Δiv for the predicted displacement, and hiv for the generated target-view representation. The target-specific parameters are denoted by θv. 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.