acceptodds
Under review as a conference paper at ICLR 2027

Coupling Error Exploration for Incomplete Multi-View Missing Multi-Label Classification

Abstract

Incomplete Multi-view Missing Multi-label Classification (IM3C) aims to predict potentially relevant labels for each sample by exploiting the consensus and complementary information across available views and labels. Although existing IM3C methods have made notable progress, the missingness of some views makes it difficult for view recovery or feature fusion to fully preserve the view-specific distributional structure, leading to the Representation Bias Problem (RBP). Furthermore, the RBP reduces the reliability of label predictions, while incomplete labels provide insufficient or even erroneous supervision, thereby interfering with representation learning and continuously amplifying prediction errors, i.e., Coupling Error Propagation (CEP). To address these challenges, we propose Coupling Error Exploration (CEE) for incomplete multi-view missing multi-label classification, aiming to achieve robust and accurate label prediction by exploiting sample-level and label-level structural information. Specifically, to alleviate RBP, we first recover the missing views and ensure the view-specific distribution structure through multi-view rectified flow learning. Furthermore, we propose multi-view structure reinforcement that utilizes multi-view contrastive learning to improve semantic consistency across views. Finally, to mitigate CEP, we design collaborative dual-graph structure learning, which enhances sample representations by leveraging sample graph relationships and explores label semantic correlations through label graph relationships. Through dual-graph collaboration, we can uncover comprehensive relational information both between samples and between views. Theoretical analyses and extensive experimental results demonstrate that CEE outperforms state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
incomplete multi-view missing multi-label classification, multi-view learning, multi-label classification, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8Db6OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses multi-view multi-label classification with both views and labels missing, and proposes CEE. A rectified flow model imputes the missing views, cross-view contrastive learning enforces semantic consistency, and a sample graph together with a label graph limits the spread of unreliable information. The framework is tested on five datasets against nine recent methods under a range of missing ratios.

Critical strengths
  1. Rather than stopping at imputing missing views, the paper also tackles error propagation through sample relationships and label dependencies, making it more complete than imputation-only or fusion-only approaches.
  2. The three modules have clear roles: MSDL recovers the view distribution, MSR enforces cross-view consistency, and CDSL handles the dual-graph learning, and they fit together naturally.
Critical weaknesses
  1. The number of neighbors, whether graphs are built within each batch or over the full training set, and the number of Euler integration steps are not stated.
  2. The classification loss is normalized by nc rather than the number of observed labels, so its relative strength varies with the label missing rate and the rationale should be explained.
  3. Pre-training updates only the velocity model and it is not stated whether the encoder is trained by then, while in joint training the encoder is updated with the generator frozen, so how latent drift is prevented needs clarification.
Questions for authors
  1. Is the encoder trained separately before rectified-flow pretraining, and how are latent space shifts handled during joint training?
  2. Please specify settings such as the number of neighbors, the scope of graph construction, and the number of ODE steps.
  3. Are the individual loss terms weighted, and why is the classification loss not normalized by the number of observed labels?
  4. How should MSDL's drops on some Pascal07 metrics be interpreted, and do the standard deviations in the tables come from the same data splits and missing masks?
1 comment

Q1 and W3: Thank you for your professional comment. The encoder is randomly initialized without separate pretraining and remains frozen in evaluation mode during the generator pretraining stage, when only the velocity model is optimized using jointly observed source-target views from the same instance. During joint training, the encoder is updated while the generator remains frozen. For each ordered pair of views that are jointly observed for the same instance, the two view representations define a linear interpolation path and a target velocity, and the generator is trained to match this velocity while being conditioned on the interpolation time and on the source and target view identities. Because the model is trained on cross-view paired instances, it learns the mapping relationships between the distributions of two or more views rather than a mapping tied to one specific coordinate system, which allows the learned transport to cope with the drift in the latent space.

Q2 and W1: Thanks for your question. The neighborhood parameters for both the sample graph and the label graph are set to 5 on all datasets. The sample graph is constructed within each training mini-batch and each evaluation batch and is rebuilt at every forward pass using the current fused representations. In contrast, the label graph is constructed once per fold from the observed labels of the entire training split and remains fixed during training; validation and test labels are excluded. The generator uses explicit Euler integration with two steps. Additional implementation settings are as follows: the contrastive projection dimension is 128, and RF pretraining uses Adam with a learning rate of 0.0002.

Q3 and W2: Thank you for drawing our attention to this distinction. All five terms in the joint objective are explicitly weighted: L=αLrec+βLcl+γLgia+σLlg+ηLcls.\mathcal{L}=\alpha\mathcal{L}_{\mathrm{rec}}+\beta\mathcal{L}_{\mathrm{cl}}+\gamma\mathcal{L}_{\mathrm{gia}}+\sigma\mathcal{L}_{\mathrm{lg}}+\eta\mathcal{L}_{\mathrm{cls}}. The classification loss is normalized by ncnc, rather than by the number of observed labels. Each observed entry contributes with weight 1/(nc)1/(nc), while missing entries contribute zero and are not treated as negative labels. Our rationale is to retain a fixed per-entry weight instead of automatically increasing the weight of the remaining observations as more labels become missing. This is a valid normalization choice, although it has a different effect from observed-label normalization. Normalization by the number of observed labels is also a reasonable alternative, and we will consider it as part of future objective-design refinements.

Q4: Thanks for your comment. Pascal07 is a complex multi-view multi-label dataset, and adding MSDL to the base model does not improve every metric, reflecting the limitations of missing-view recovery alone in this setting. The components of our method are mutually reinforcing. As MSR and CDSL are subsequently incorporated, the model achieves substantially better performance, benefiting from the combined effects of missing-view recovery, cross-view consistency, and relational learning. The standard deviations in the tables come from the same data splits and missing masks.

Reviewer_8a9DOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes the Coupling Error Exploration (CEE) method, designed to address the Representation Bias Problem (RBP) and the issue of Coupling Error Propagation (CEP) within the context of incomplete multi-view missing multi-label classification. The method employs multi-view rectified flow learning to recover missing view features within the latent space. Following the completion of the multi-view data, the authors utilize multi-view contrastive learning to enhance the semantic consistency across views, while constructing view-graph and label-graph relationships to explore the underlying connections within the data. The experiments evaluate the method’s performance under various missing-data conditions, with results demonstrating that the approach consistently maintains a significant advantage across these diverse scenarios.

Critical strengths
  1. The manuscript introduces a rectification-based learning strategy to model distribution transformations from source to target views in the latent space. By integrating predictions generated from multiple observed views, it mitigates representation bias in missing-view recovery.
  2. The manuscript develops a collaborative dual-graph structure learning framework. Information compression and adaptive gating are incorporated into instance-graph learning to suppress unreliable neighborhood information. Meanwhile, label observation patterns inform label correlation estimation and regularize classifier learning, improving the reliability of feature learning and label prediction.
  3. The manuscript presents extensive experiments that provide empirical support for its theoretical claims and demonstrate the effectiveness of the proposed method.
Critical weaknesses
  1. Certain implementation settings are not explicitly specified, such as the number of Euler integration steps and the neighborhood sizes for the sample graph and label graph. Providing this information would help improve the reproducibility of the research. 2.The framework comprises several learning modules, yet it lacks an analysis of computational efficiency. Including an analysis of time complexity would allow readers to evaluate its computational overhead and potential for practical application.
  2. The modular integration method used in the ablation study is additive, which makes it difficult to distinguish the independent contribution and synergistic effect of each module. Adding control experiments by removing MSDL, MSR, and CDSL from the complete model separately to more clearly verify the role of each module. 4.A more detailed discussion of the limitations of the proposed method would help make the assessment of CEE more balanced and comprehensive.
Questions for authors
  1. Is the sample graph built within each minibatch, over the training set, or over evaluation batches, and what is the neighbor count and update schedule?

  2. In Eq. (13), the squared distance is symmetric, so the penalty depends only on (A_L + A_L^T)/2 and is unchanged when A_L is transposed. What part of the model exploits the direction of the estimated label relations?

1 comment

Q1 and W1: Thanks for your question. The sample graph is constructed within each training mini-batch and each evaluation batch during validation and testing. We select the five nearest neighbors for each sample. It is rebuilt at every forward pass using the current fused representations. The generator uses Euler integration with two steps. The number of neighbors for both the sample graph and the label graph is set to 5.

Q2: Thanks for your question. Directionality is incorporated when estimating conditional label co-occurrence probabilities and selecting label-specific neighbors during graph construction. Although the estimated conditional label relations can be asymmetric, Eq.(13) does not explicitly exploit their direction; it uses the resulting pairwise association strengths to encourage consistency between the classifier weights of related labels.

W2: Thank you for the suggestion. We have added a time complexity analysis to clarify the computational cost and scalability of our method. Specifically, we analyze the complexity of each loss function, the encoder and decoder, and the construction of the sample and label graphs, and then summarize their combined computational cost. The specific content is presented as follows:

Complexity Analysis. Let Dmax⁡D_{\max} denote the maximum number of neurons in the intermediate network layers. The computational complexities of Lrf\mathcal{L}_{\mathrm{rf}}, Lrec\mathcal{L}_{\mathrm{rec}}, Lcls\mathcal{L}_{\mathrm{cls}}, Llg\mathcal{L}_{\mathrm{lg}}, Lgib\mathcal{L}_{\mathrm{gib}}, and Lcl\mathcal{L}_{\mathrm{cl}} are O(nm2Dmax⁡)O(nm^{2}D_{\max}), O(nmDmax⁡)O(nmD_{\max}), O(nc)O(nc), O(nc2Dmax⁡)O(nc^{2}D_{\max}), O(nDmax⁡)O(nD_{\max}), and O(nm2Dmax⁡)O(nm^{2}D_{\max}), respectively. The computational complexities of the encoder and decoder are O(nmDmax⁡2)O(nmD_{\max}^{2}). The computational complexities of Sample-graph and label-graph construction are O(nDmax⁡)O(nD_{\max}) and O(nc2)O(nc^{2}). The combined time complexity of these components is O(nm2Dmax⁡+nmDmax⁡+ncO(nm^{2}D_{\max}+nmD_{\max}+nc +nc2Dmax⁡+nDmax⁡+nc^{2}D_{\max}+nD_{\max} $+nmD_{\max}^{2}+nD_{\max}+$$nc^{2})$. In summary, the analyzed computational cost depends primarily on the network width, the number of views, and the number of labels. The analyzed cost scales linearly with the number of samples nn, which demonstrates the good scalability of our framework in multi-view scenarios.

W3: Thanks for your comment. We conducted non-cumulative ablations on Corel5k and Pascal07 with both view and label missing rates set to 50%. Starting from the full model, we separately disable the MSDL, MSR, and CDSL. The details are as follows:

Corel5kPascal07
MethodAP ↑1-RL ↑AUC ↑1-OE ↑AP ↑1-RL ↑AUC ↑1-OE ↑
CEE w/o MSDL0.39540.90850.91170.45170.55740.83070.85350.4700
CEE w/o MSR0.41440.90600.90930.49180.55520.82880.85280.4693
CEE w/o CDSL0.40900.90720.91020.46820.55610.83620.85780.4691
CEE0.43250.91700.91940.50080.56410.83920.86080.4782

The experimental results indicate that the absence of either MSDL, MSR, or CDSL resulted in a significant decline in model performance. The experimental results lead to the following conclusions: (1) MSDL improves AP by 3.71 and 0.67 percentage points on Corel5k and Pascal07, respectively. These gains support its role in recovering missing-view representations through view-specific distribution learning. (2) MSR improves AP by 1.81 and 0.89 percentage points, respectively, benefiting from stronger cross-view consensus that improves representation consistency and discriminative learning. (3) CDSL improves 1-OE by 3.26 and 0.91 percentage points, respectively, indicating better identification of relevant labels. This improvement benefits from combining sample-level neighborhood aggregation with label-dependency regularization, enriching sample representations while constraining classifier learning.

W4: Thank you for your suggestions. We have added a discussion regarding the limitations of CEE, as follows:

Limitations. Although our method performs well on the evaluated benchmarks, it has several limitations. First, under high view-missing rates, fewer samples retain jointly observed source and target views, reducing the paired supervision available for generator training. This may limit the recovery of missing-view information. Second, computational costs can become substantial when the number of views or labels is large. More views increase the cost of cross-view learning, while larger label spaces increase the cost of label-relation estimation and graph regularization. Future work will explore more data-efficient generation strategies and computationally efficient extensions for larger-scale applications.

Reviewer_iznJOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies incomplete multi-view missing multi-label classification (IM3C), which involves scenarios where both views and labels are simultaneously missing. The authors introduce two challenges in IM3C: Representation Bias Problem (RBP) from missing views,and Coupling Error Propagation (CEP) between unreliable predictions and incomplete supervision. To address these challenges, the authors propose the Coupling Error Exploration (CEE) framework. CEE recovers representations of missing views from observed views by learning a velocity model within the latent space. The authors ensure semantic consistency through multi-view contrastive learning and facilitates relational learning by constructing both sample graphs and label graphs. The authors evaluate CEE against nine recent baseline models across five multi-view multi-label datasets, utilizing six label-related evaluation metrics. The experiments assess the performance of CEE under varying rates of missing views and labels, and include an ablation study of the framework's three core components. The results demonstrate that CEE maintains a robust performance advantage even as the missingness rates increase.

Critical strengths
  1. This manuscript examines the impact of simultaneous view and label incompleteness on multi-view multi-label classification from a novel perspective termed Coupling Error Exploration.

  2. This manuscript selectively incorporates neighborhood information into the instance graph through a Gaussian information bottleneck and a gating mechanism. Meanwhile, the label graph estimates label co-occurrence relationships by accounting for annotation availability. Together, these components compensate for information loss caused by missing views and labels at both the instance and label levels.

  3. This manuscript derives a variational upper bound on mutual information in neighborhood representation learning and a tractable Gaussian KL divergence formulation. The accompanying proofs and analysis provide theoretical justification for the proposed mechanism and clarify its underlying rationale.

  4. Extensive experiments across multiple datasets and comparisons with diverse baselines substantiate the effectiveness of the proposed method. The complementary theoretical analysis and empirical evaluation strengthen its methodological rigor.

Critical weaknesses
  1. The generator is trained with fixed encoders and then frozen while the encoders are updated; how its learned transport remains compatible with the changing latent spaces is unspecified.

  2. The existing missingness sweeps demonstrate useful predictive robustness, but do not isolate propagation through corrupted sample relations or unreliable label dependencies.

Questions for authors
  1. How are the autoencoders initialized or pretrained before the 400-epoch generator stage, when all parameters except the velocity model are fixed ?

  2. Could the authors directly evaluate recovery on held-out observed views that are artificially masked, using both paired reconstruction and distributional discrepancies ?

  3. Distinguish marginal-distribution recovery from instance-specific reconstruction and explain why averaging in Eq. (6) should retain target-view structure rather than blur differing predictions.

  4. Could the authors separate the sample graph, label graph, Gaussian compression and residual gate in the ablation, and remove MSDL from the otherwise full model ?

2 comments

Q1 and W1: Thank you for your comment. The autoencoders are randomly initialized without separate pretraining and remain frozen during the 400-epoch generator stage. For each ordered pair of views that are jointly observed for the same instance, the two view representations define a linear interpolation path and a target velocity, and the generator is trained to match this velocity while being conditioned on the interpolation time and on the source and target view identities. Because the model is trained on cross-view paired instances, it learns the mapping relationships between the distributions of two or more views rather than a mapping tied to one specific coordinate system, which keeps the learned transport compatible with variations in the latent space.

Q2: Thank you for this suggestion. We evaluated view recovery on Pascal07 by comparing the generated representations with the corresponding ground-truth representations of artificially masked views. The details are as follows:

ViewAIMNetOurs
Cosine ↑Euclidean ↓Cosine ↑Euclidean ↓
View 10.262013.54840.313413.3464
View 20.085420.96380.288320.1388
View 30.075818.26200.165718.1192
View 40.60488.82910.62918.1574
View 50.86155.40170.80444.7350
View 60.79493.84710.75233.6889

Overall, the consistently lower Euclidean distances and improved cosine similarity on most views support the effectiveness of our generator in recovering missing-view representations across different target views. These results provide evidence of its cross-view recovery capability under the evaluated setting.

Q3: Thank you for raising this insightful question. Eq. (6) preserves the target-view representation when it is observed and averages predictions from the available source views of the same instance only when the target view is missing. Our generator uses paired source–target representations from the same instance for training, providing instance-level supervision rather than only marginal matching. Averaging therefore combines predictions for the same target representation and is intended to reduce source-specific errors, rather than mixing different target-view structures.

Q4:Thanks for your question. We have conducted non-cumulative ablations on Corel5k and Pascal07 with both view and label missing rates set to 50%. The details are as follows:

Corel5kPascal07
MethodAP ↑1-OE ↑AP ↑1-OE ↑
CEE w/o MSDL0.39540.45170.55740.4700
CEE w/o Sample Graph0.42690.49930.55770.4723
CEE w/o Label Graph0.41270.49310.55760.4709
CEE w/o Gaussian Compression0.43190.49490.56000.4718
CEE w/o Adaptive Residual Gate0.42570.49500.56180.4720
CEE0.43250.50080.56410.4782

The experimental results support the following contributions: (1) MSDL improves 1-OE on Corel5k through view-specific distribution learning that recovers missing-view representations. (2) The sample graph improves AP on Pascal07 by aggregating complementary neighborhood information to enrich sample representations. (3) The label graph improves AP on Corel5k by incorporating label dependencies into classifier learning. (4) Gaussian compression improves 1-OE on Pascal07 by regularizing the information retained in aggregated representations. (5) Adaptive residual gating improves AP on Corel5k by adaptively balancing original features and neighborhood information.

W2: Thanks for your comment. To directly assess the downstream effects of damaged sample relations and disrupted label dependencies, we evaluated this by varying the label-missing rate while fixing the view-missing rate at 50%, and by varying the view-missing rate while fixing the label-missing rate at 50%. The details are as follows:

View missing rate fixed at 50%Label missing rate fixed at 50%
Missing rateMethodAP ↑AP ↑
0.0DIMC0.38200.4511
LMVCAT0.41250.4743
Ours0.45260.4858
0.3DIMC0.3701 (↓0.0119)0.4114 (↓0.0397)
LMVCAT0.3995 (↓0.0130)0.4201 (↓0.0542)
Ours0.4509 (↓0.0017)0.4645 (↓0.0213)
0.5DIMC0.3544 (↓0.0276)0.3544 (↓0.0967)
LMVCAT0.3666 (↓0.0459)0.3666 (↓0.1077)
Ours0.4325 (↓0.0201)0.4325 (↓0.0533)
0.7DIMC0.3329 (↓0.0491)0.2797 (↓0.1714)
LMVCAT0.3453 (↓0.0672)0.3185 (↓0.1558)
Ours0.4155 (↓0.0371)0.3441 (↓0.1417)

Our method achieves the highest AP and the smallest absolute AP decreases across all tested nonzero missing rates in both settings. These results support that our method is robust against error propagation resulting from compromised samples or reliance on unreliable labels.

Thank you to the authors for the detailed responses and additional experiments, which have addressed my concerns. My score remains unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.