Class-Incremental Continual Learning for Multi-View Clustering

1. The proposed method only learns the common feature sapce, and ignores the contribution of heterogeneity among multiple views to the clustering.
2. Although the experimental results demonstrate the effectiveness of the proposed method, the two-stage independent optimization approach carries the risk of inconsistent optimization.
3. The paper lacks a computational complexity analysis. With an increasing number of views, is the clustering performance and the computational complexity of the proposed method stable?

W1. Although the paper provides a well-structured exposition of the proposed methodology, further technical insights regarding the implementation and specific algorithms within the CCMVC method would be beneficial.

W2. The manuscript could delve deeper into the significance of class-incremental multi-view clustering for real-world applications, a factor which could be potential to inspire future research in this domain.

W3. The paper falls short in providing a detailed analysis of the limitations of the proposed approach, a factor which could be significant for future research and practical applications.

W4. The computational complexity of the proposed algorithm, which could be a concern for large-scale datasets or real-world applications, is not discussed in the manuscript.

W5. A more detailed exposition of the datasets used in the evaluation, including their characteristics and potential biases, would enrich the manuscript.

1. The main concern is regarding the experiment. The dataset used is relatively small (100010000 samples with 710 classes) to effectively validate the effectiveness of the proposed method. Considering the problem revolves around increasing classes, the inclusion of only about 2 new classes may not sufficiently demonstrate the ability of the proposed method in handling new classes.
2. The approach of cross-view contrastive learning seems not novel to the community, as it is largely based on contrastive clustering methods with the only difference being the contrast of clustering anchors.
3. The complexity of the proposed method is not adequately discussed. It would be helpful to compare the computation cost of the proposed method to the baselines.
4. The choice of using GMM for clustering is not justified. It would be beneficial if the authors provided a reason for this selection or discussed alternative methods.

1. In INTRODUCTION, the authors propose two challenges. However, I think they can be summarized as one challenge.
2. In NOTATION AND PROBLEM DEFINITION, it is obtained that each task has the same sample number $N$. It is more rational to set the sample number as $N_t$.
3. The method randomly selects samples to add to the buffer. The clustering performance might be sensitive to the selected samples. Also, since it is a random sampling strategy, the sample number in each cluster is not balanced. In extreme circumstances, there might be no samples in a cluster. I wonder if it affects the performance of the method.
4. The Gaussian mixture models and parameters share among views. MVC aims to extract consistent and complementary information among views. The assumption that the data of different views have a similar distribution might be unsuitable.
5. Some recent works could be discussed, such as [1].
6. The used datasets are small, and the time efficiency could be compared. I wonder if the method can address large-scale data. [1] Li, D., Wang, T., Chen, J., Kawaguchi, K., Lian, C., & Zeng, Z. (2023). Multi-View Class Incremental Learning. ArXiv, abs/2306.09675.