Masked Two-channel Decoupling Framework for Incomplete Multi-view Weak Multi-label Learning

Multi-view learning has become a popular research topic in recent years, but research on the cross-application of classic multi-label classification and multi-view learning is still in its early stages. In this paper, we focus on the complex yet highly realistic task of incomplete multi-view weak multi-label learning and propose a masked two-channel decoupling framework based on deep neural networks to solve this problem. The core innovation of our method lies in decoupling the single-channel view-level representation, which is common in deep multi-view learning methods, into a shared representation and a view-proprietary representation. We also design a cross-channel contrastive loss to enhance the semantic property of the two channels. Additionally, we exploit supervised information to design a label-guided graph regularization loss, helping the extracted embedding features preserve the geometric structure among samples. Inspired by the success of masking mechanisms in image and text analysis, we develop a random fragment masking strategy for vector features to improve the learning ability of encoders. Finally, it is important to emphasize that our model is fully adaptable to arbitrary view and label absences while also performing well on the ideal full data. We have conducted sufficient and convincing experiments to confirm the effectiveness and advancement of our model.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer XPPT6/10 · confidence 5/52023-06-27

Summary

The core innovation of the method lies in decoupling the single-channel view-level representation, which is common in deep multi-view learning methods, into a shared representation and a view-proprietary representation with a cross-channel contrastive loss. The authors have conducted sufficient experiments to verify the effectiveness of the proposed method.

Strengths

1. The definition of the problem in this paper is clear. The topic is novel, and as far as I know, multi-view multi-label classification methods for the complex situation of missing labels and data are underexplored. 2. The interpretation of extracting multi-view features from two perspectives is intuitive and fits with the fundamental assumption of multi-view/multi-modal learning problems that there are commonalities as well as differences between modalities. 3. The experimental results are sufficient. The authors not only provide the results in the missing case, but also provide the results of the proposed model on the complete datasets, which shows that the model has good learning ability for multi-view data.

Weaknesses

1. In Eq.(1), the necessary explanation about the coefficient 2 in the numerator is lacking. The authors should also provide more explanation about the motivation of such design. 2. A clarification on the value of mask rate $\sigma$ is lacking. 3. The authors propose to build two groups of encoder-decoder for different views, but the computational complexity may increase with the number of views quickly. 4. There are some unclear expressions, for instance, in Section 2.2, the authors say “Different from conventional deep multi-view networks”. So what are the conventional deep multi-view networks? There is no corresponding reference here.

Questions

Is it efficient to address the missing problem by introducing additional prior information, that is matrix G and W in this paper?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have discussed the limitations and suggested further directions for improvement in the Conclusion.

Reviewer txHg6/10 · confidence 5/52023-06-30

Summary

This paper proposes a general framework for missing multi-view and missing multi-label classification. The main innovation is to decouple each view feature into two different shared and private features. The paper also applies random masks to the raw features to make the network learn from limited information, and shows the benefits of this approach with ablation experiments.

Strengths

- This paper is clear in terms of writing and presentation. It is easy for the reader to understand the author's motivation. - The problem addressed by the authors is complex (doubly incomplete data) and poses a greater challenge than methods based on ideally complete data. However, judging from the results, the author solves this problem well. - The method of using labels to constrain the feature extraction is novel, and the author's local masking of input features has also brought performance growth. I think there are simple and effective techniques that can be extended to other fields. - The experimental results show that the MTD has significant advantages over the existing multi-view multi-label classification methods.

Weaknesses

- The one-HL metric does not seem to be of much significance because it does not effectively evaluate the performance of different methods according to the author's experimental results. - It is inappropriate to appear both ‘Figure’ and ‘Fig.’ in the text. - The author should clearly explain the meaning of Figure 3 in the caption.

Questions

- Are these masks fixed during training? How exactly did the authors set up the mask? - Why the results of iMVWL and NAIM3L are not consensus with original papers?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer 8GYm6/10 · confidence 4/52023-07-03

Summary

This paper proposes a new framework for incomplete multi-view weak multi-label classification (iMvWMLC), which is a challenging task that involves missing views and labels in the data. The framework consists of four main components: a two-channel decoupling mechanism (called MTD) that extracts shared and view-specific features from each view, a cross-channel contrastive loss that enhances the semantic separation of the two channels, a random fragment masking strategy that reduces the redundancy of the input features, and a label-guided graph regularization loss that preserves the geometric structure among samples. The paper evaluates the framework on five datasets and compares it with eight state-of-the-art methods, showing that it outperforms them on various metrics and is robust to different levels of incompleteness.

Strengths

(1) As far as I know, research on multi-view multi-label classification is still in its infancy, and this article focuses on this topic very well. (2) To a certain extent, the logic and expression of this article are clear. (3) I am interested in masking operations applied to vectorized feature data, most multimodal methods focus more on input real data such as text or images. I agree that the use of the mask mechanism on feature data is instructive.

Weaknesses

(1) The similarity matrix calculated based on weak tags does not take into account the interference caused by unknown tags. The paper should address the potential interference caused by unknown tags in the calculation of the similarity matrix. (2) The results of the ablation experiment suggest that the so-called contrast loss does not seem to have improved much, at least on the Corel5k dataset. The paper should provide some explanation. (3) The paper needs to improve the details, as there is a error in section 3.4 where "TMD" is incorrectly referenced. The authors could correct this and prevent similar errors. (4) The reviewer has noticed that there are some recent works aiming at solving the multi-view classification problem. The paper could do a better job by citing the different relevant works, e.g., [1-3], and clarifying the differences between this work and them. [1] Trusted Multi-View Classification, ICLR’21 [2] Partially View-aligned Representation Learning with Noise-robust Contrastive Loss, CVPR’21 [3] Dual Contrastive Prediction for Incomplete Multi-View Representation Learning, TPAMI'23

Questions

(1) It would be interesting to know if the authors have experimented with other masking methods, such as random bit masking of features, and what the results of such experiments were. (2) Availability of code is important for reproducibility and further research. There is no code in the supplement material. Do the authors have any plans to make the code public? (3) The paper should provide a more detailed discussion on why not all comparison methods were adapted to this task.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

For missing multi-view data without a priori missing information, this method seems to be difficult to work.

Reviewer NrND4/10 · confidence 4/52023-07-12

Summary

This paper studies incomplete multi-view weak multi-label learning problem, which is important. The authors propose a masked two-channel decoupling framework based on deep neural networks. They develop cross-channel contrastive loss, a label- guided graph regularization loss, and random fragment masking strategy. The experiments validate the effectiveness.

Strengths

- The problem is interesting. - The techniques in this work seem correct and reasonable.

Weaknesses

- This work just combines some existing widely-used techniques. Thus the work is a bit incremental, does not provide new insights to me. - As stated in abstract, the core is to decouple view-level representation into shared representation and a view-proprietary representation. This idea has been widely used in existing works. - The authors first fill the missing views with noise, and missing label with 0, and then treat it as normal complete view learning problem. To me, this scheme seems not grasp the essence of multi-view weak multi-label learning problem. - The aim and construction of contrastive loss in this work seem a bit confusing to me. It is suggested to explain it in more detail.

Questions

See weakness part

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The author seems to lack a comprehensive discussion on the limitations of the proposed method.

Authorsrebuttal2023-08-10

Response to weakness 4

W4. The aim and construction of contrastive loss in this work seem a bit confusing to me. It is suggested to explain it in more detail. A4. Thanks for the comment, and we are sorry to confuse you about the contrastive loss. According to your suggestion, we further explain the design idea of contrastive loss as follows in the manuscript: For any sample $i$, we pair features on $2N$ channels in pairs ($4N^2$ feature pairs in total), and classify these feature pairs into 2 categories, namely positive pairs and negative pairs, where positive pairs consist of shared features from different views, and the remaining “shared-private” and “private-private” feature pairs are negative pairs. Our goal is to minimize the distance between positive instance pairs (the denominator part), while maximizing the distance between negative pairs (the numerator part). Here we not only consider maximizing the difference between private features decoupled from the different view ($o_{i}^{(u)}$ and $o_{i}^{(v)}$, $u\neq v$), but also reduce the similarity of shared features and private features from both intra-view and inter-view levels ($s_{i}^{(u)}$ and $o_{i}^{(v)}$, including $u=v$). This approach fulfills the design objective of the dual-channel model, which is to encourage consistency between shared features from different views while maintaining a clear distinction between view-proprietary feature and shared or other view-proprietary features. In addition, considering the efficiency of matrix calculation, we set the coefficient of the contrastive item of shared features and private features to 2 in the numerator part, so that the entire paired similarities can be obtained by a single matrix multiplication, i.e., $[s_i^{(1)}; s_i^{(2)};…; s_i^{(N)}; o_i^{(1)}; o_i^{(2)};…; o_i^{(N)}]\times [s_i^{(1)}; s_i^{(2)};…; s_i^{(N)}; o_i^{(1)}; o_i^{(2)};…; o_i^{(N)}]=Q\in \mathbb{R}^{2N\times 2N}$. The matrix $Q$ can be divided into four sub-matrices whose size is $N$ by $N$. The sub-matrix in the upper left corner stores the similarity values of the denominator in Eq. (1), the sub-matrix in the lower right corner stores the values of the second item in the numerator, and the remaining two submatrix, which are transposed to each other, store the similarity values of the first item in the numerator. Therefore, we set the coefficient of the first term in the numerator to 2. At the same time, the diagonal elements of $Q$ are removed, so that the number of elements in the numerator and denominator parts is $\frac{1}{3N^{2}-N}$ and $\frac{1}{N^{2}-N}$ respectively. About limitation of the paper: In the conclusion, we explain the limitations of our method and further improvement directions in the future, such as adding the consideration of label correlation to the multi-label classification problem in the model. At the same time, for missing views and partial multi-label, in addition to introducing prior information, more complex view completion and pseudo-label prediction can be performed to improve the discriminability of the model.

Reviewer NrND2023-08-22

Thank you for your feedback. The authors have addressed some of my concerns and I will increase my score.

Reviewer XPPT2023-08-12

Thanks for the authors' responses. The authors have addressed my concerns and I'd like to rise my score.

Authorsrebuttal2023-08-21

Response to Reviewer XPPT

Thank you very much for your review comments and encouragement, we will further improve the quality of the manuscript.

Reviewer 8GYm2023-08-14

Thanks for the authors' responses. The authors have addressed my concerns and I will maintain my rating of acceptance for the paper.

Reviewer txHg2023-08-20

Thank you for the responses. The responses have solved all my concerns. I am satisfied with it and would like to keep my previous scores (weak accept).

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC