Efficient Incomplete Multi-view Clustering via Flexible Anchor Learning

1.This paper is a modification of previous work published in CVPR (https://ieeexplore.ieee.org/document/9880247). It lacks any novel insights into incomplete multi-view anchor graph clustering, which limits its contribution. Based on this, I do not believe it is suitable for acceptance at ICLR.

2.The construction of the incomplete anchor graph follows the approach in the aforementioned work, and the graph partition method (matrix factorization) is also a similar method in TIP-21(https://ieeexplore.ieee.org/abstract/document/9305974). The paper does not offer any significant new ideas or contributions.

3.Incomplete anchor graph clustering is not a novel topic and has been extensively studied. Consequently, the significance of this paper's contribution appears minimal.

4.The paper does not compare its results with any deep learning-based baselines for incomplete data, which limits the assessment of its effectiveness.

5.The optimization method used is alternative optimization, no new techniques or improvements are proposed.

6.The experiments conducted are quite common and fail to deliver any particularly exciting results.

1. The authors state that each entry in the anchor graph $Z$ describes the similarity between data sample and anchor. Since the symmetric constraint on $Z\in R^{m\times n}$ are not guaranteed in factorization with $m\ll n$, the authors remove such constraint on anchor graph $Z$ and this is the main difference between anchor graph and similarity matrix in symmetric nonnegative matrix factorization. In this part, the authors are expected to give the reason why each entry in the anchor graph $Z$ describes the similarity between data sample and anchor.

2. The authors list the detailed clustering results of EIML and the compared approaches on different datasets in terms of four metrics in Tables 1-4. The authors also compare EIML with IMVC-CBG and FIMVC-VIA under different missing ratios on several datasets under different metrics. According to Tables 1-4 and Figs. 4-7, the authors then draw some following conclusions. However, the authors do not bold the best clustering performance in Tables 1-4 in terms of four metrics for this paper.

3. The authors perform the parameter selection for the trade-off parameter $\lambda$ in the range of $[0.001,0.1,1,10,100,1000]$ to study how these this parameter influences the final clustering performance and find that better performance is achieved when $\lambda=1$ under the same $m$ on different datasets. Besides, the clustering result of EIML is relatively stable over different parameter values on these datasets, which shows that EIML is generally robust to the trade-off parameter $\lambda$. However, the authors do not give the detailed reason why the $[0.001,0.1,1,10,100,1000]$ is chosen for parameter selection in this paper.

4. The authors perform ablation study to validate the superiority of adopting a unified framework integrated by graph construction, anchor learning and graph partition. In comparative experiments, the authors first learn anchors and construct the graph to obtain informative representation. Then the graph partition is isolated from the above two processes in the designed experiment. However, the authors do not give the detailed specific values analysis regarding the superiority of adopting a unified framework integrated by graph construction, anchor learning and graph partition.

1. This paper lacks novelty. In fact, anchor graph learning and symmetric nonnegative matrix factorization are two very common methods in multi-view clustering. This paper is an easy combination of these two methods. I think that the quality of this paper does not meet the standards of ICLR.
2. Since the objective function includes the cluster assignment matrix $F$, I think that the initialization may heavily influence the clustering results. The authors should conduct experiments to verify whether the different initializations have different clustering performances.
3. The paper lacks the necessary theoretical explanations, such as a convergence analysis.
4. The resolution of Figure 1 is too low and needs to be adjusted.

The proposed method does not address missing views in a meaningful way, lacking specific techniques for incomplete multi-view clustering. This deviation from the core theme reduces the relevance of the approach to the stated problem of incomplete multi-view data.

The experimental section lacks sufficient comparison methods, with limited engagement with recent advancements in the field. Additionally, the description of datasets is brief and does not adequately highlight the method’s capacity to handle large-scale data.

The illustrations in the paper are of low quality, lacking clarity and depth. They do not effectively convey the core concepts, which may hinder the reader’s understanding of the proposed method.

The layout of the paper is inconsistent, with poor alignment between text descriptions and related figures or tables. This inconsistency disrupts the reading flow and detracts from the paper's overall presentation quality.