- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes a novel framework for trustworthy multi-modal clustering. It integrates information bottleneck theory with evidential deep learning, which represents a fresh and reasonable direction that has not been fully explored before. The experiments are sufficient and cover multiple public multi-modal datasets. The evaluation is conducted on several benchmark datasets and compared against a broad set of recent baselines, which lends credibility to the reported results.
Critical strengths
- The paper addresses a relevant problem in unsupervised Earth observation. Unreliable neighborhood relationships and differences across sensing modalities are practical challenges, and the proposed framework tackles both through a unified training objective.
- The evaluation covers four scenes and includes comparisons with the closely related GMAE baseline, component ablations, and controlled experiments examining the effects of data corruption and geometry. The paper also reports using ten random seeds and paired initialization for the alignment comparisons.
- The paper is well motivated, with the main challenges clearly linked to the proposed components. The experimental figures are easy to follow, and the qualitative clustering maps help illustrate differences in spatial coherence and boundary preservation across methods.
Critical weaknesses
- The parameter range stated in Section 3.2 is inconsistent with the experiments in Figure 10. Section 3.2 requires , whereas Figure 10 includes , , and , without explaining why values outside this range are tested. In particular, makes the reconstruction loss constant wherever the power is defined, so it no longer provides a useful reconstruction signal for training.
- The role of the graph learned by DGPO in the overall training pipeline is not clearly explained, particularly its connection to the graph encoder.
- Appendix A.3 lists many dataset-specific parameter settings but does not explain how they were selected. It is also unclear whether the baselines used comparable input features and tuning budgets.
- The initial and purified graphs on Salinas in Figures 3(a) and (b) look fairly similar. Although purification reduces some off-diagonal connections, the improvement is less evident than the description in Section 4.5 suggests, particularly regarding the formation of a more cohesive block-diagonal structure and the suppression of misleading edges.
- Some references in the Appendix are not properly rendered. For example, “Figure ??” appears in Appendix C.2, in the subsection titled “Sensitivity to the entropy regularization coefficient .” The citation below Eq. (102) in the proof of Theorem 8 also appears as “?”.
Questions for authors
- Could the authors clarify the intended range of gamma and explain why Figure 10 includes = 0, 0.1, and 0.5 despite the restriction ≥ 1 in Section 3.2? How is the reconstruction loss handled at = 0, and does this setting effectively remove its contribution to model training?
- Could the authors explain how the graph learned by DGPO is used during training, particularly its connection to the graph encoder? Is the updated graph used for message passing in later training iterations, or does it affect training through the structural and topology-consistent reconstruction losses?
- Please describe how the dataset-specific parameter settings in Appendix A.3 were selected? Please also specify the input features and preprocessing used for the baselines, and explain how their parameter-tuning budgets compare with that of AM-GMAE.
- In Figures 3(a) and (b), the most visible change is a reduction in some off-diagonal connections, while the overall block structure appears similar. Could the authors explain how this comparison supports the claims in Section 4.5 about a more cohesive block-diagonal structure and the suppression of misleading edges? A clearer visualization highlighting the relevant changes would be helpful.
7 comments
W3 & Q3: Thank you for asking us to detail the hyperparameter selection strategy, baseline preprocessing, and tuning fairness. For AM-GMAE, general training configurations including the learning rate (), training epochs , masking ratio , reconstruction exponent , and initial threshold are fixed across all datasets. The dataset-specific parameters in Appendix A.3, namely the loss weights , temperatures , cross-view saliency weight , and Sinkhorn regularization , were selected via grid search. Specifically, we searched the loss coefficients over and the remaining parameters over the candidate grids shown in the sensitivity analyses of Appendix C.2 (Figures 8 to 13), and then reported the mean performance over 10 random seeds .
For all baseline methods, we use the same multi-view input modalities, feature dimensions, and feature normalization as AM-GMAE across all four datasets, as described in Table 3 and Appendix A.1. For superpixel-based remote sensing graph baselines such as SAMVGC and CG-MAE, we provide the same SLIC superpixel segmentation, region-averaged node features, and initial -NN graphs used by our method. For the remaining baselines, we follow their official preprocessing and graph or anchor construction pipelines on the same normalized multi-view features.
To ensure a fair parameter-tuning budget across all methods, we evaluated each baseline using its official open-source code under a consistent protocol. For baselines whose original papers already reported experiments on the corresponding dataset, we directly adopted the optimal hyperparameters provided by the authors. For baselines that had not been evaluated on a given dataset in their original papers, we first tested their default parameter settings and then conducted a grid search over their key hyperparameters within the candidate ranges recommended by the original authors to explore better performance. Finally, using each baseline's best-performing parameter configuration, we ran 10 independent trials with the same random seeds and reported the average results in Table 1.
W4 & Q4: Thank you for this helpful suggestion regarding the visualization in Figure 3. In the original submission, Figures 3(a) and 3(b) were plotted on the Salinas dataset, which has 16 fine-grained classes and 3,500 superpixel nodes; as a result, the individual class blocks were relatively small, and the visual changes were less obvious at full-matrix resolution. In the revised manuscript (Section 4.5, page 9), we have redrawn Figures 3(a) and 3(b) on the Trento dataset and enhanced the visual contrast of the adjacency matrices. In the updated figures, the initial graph in Figure 3(a) exhibits substantial off-diagonal inter-class noise, whereas the DGPO-purified graph in Figure 3(b) shows a much cleaner and more cohesive block-diagonal structure with off-diagonal connections clearly suppressed.
Furthermore, the claims in Section 4.5 are quantitatively supported across all four datasets by the structural metrics in Figure 3(c) and the inter-class edge weight distributions in Figure 3(d) (detailed in Appendix A.3). On Trento, DGPO improves edge purity from to , weighted homophily from to , modularity from to , and increases the intra/inter-class edge weight ratio from to . Similarly, on Salinas, DGPO increases edge purity from to , weighted homophily from to , modularity from to , and more than doubles the intra/inter-class weight ratio from to . Consistently, the KDE curves in Figure 3(d) show that DGPO shifts the distribution of remaining inter-class edge weights toward lower values across all four datasets, confirming that misleading cross-class edges are effectively suppressed.
W5: Thank you for catching these rendering errors. They were caused by typographical mistakes in the LaTeX labels and BibTeX keys during compilation, and we have corrected all of them in the revised manuscript. Specifically, the broken reference in Appendix C.2 under the subsection on the entropy regularization coefficient now properly renders as Figure 11 alongside the corresponding plot on page 36, and the missing citation below Eq. (102) in the proof of Theorem 8 now properly renders as (Benamou et al., 2015). We have also checked the entire manuscript and appendix to ensure that all cross-references and citations compile properly.
We sincerely thank you for the constructive feedback and insightful questions. We address your specific concerns in detail below.
W1 & Q1: Thank you for pointing out this inconsistency and giving us the opportunity to clarify the range of γ values. In Section 3.2, the condition γ≥1 defines the standard operating range of the scaled cosine reconstruction loss during regular training. In all main experiments, we set γ=1. In contrast, Figure 10 in Appendix C.2 is intended to provide a broader sensitivity analysis; therefore, we initially tested values between 0 and 1 to observe how clustering performance changes as the reconstruction penalty is weakened. To avoid misunderstanding and fully cover the case where γ≥1, we have expanded Figure 10 in the revised manuscript to include γ∈{0.0,0.1,0.5,1.0,2.0,3.0,5.0}.
When γ=0, the 0-th power of the cosine reconstruction error becomes the constant 1, which results in a zero gradient, thereby rendering LRecon completely ineffective during training. We treat γ=0 as a boundary case in our ablation experiments, aiming to evaluate the model’s performance when “direct masked attribute reconstruction” is disabled. Under this setting, the model does not collapse because the remaining parts of the objective function in Equation (13) still provide gradients for updating the encoder and decoder. However, as shown in Figure 10, removing LRecon when γ=0 leads to a decrease in the model’s accuracy across all datasets; this confirms that introducing an effective reconstruction loss is essential for achieving optimal clustering performance.
W2 & Q2: Thank you for asking us to clarify how the purified graph A~(v) learned by DGPO is used during training. We have updated Appendix A.4 in the revised manuscript to present the complete training flow in detail. In each training epoch, the forward pass of the graph encoder always uses the normalized initial k-NN graph A~(v) together with the CSAM-masked node features to compute the view-specific latent representations H(v), and the decoder maps H(v) to the reconstructed node features Z^(v). Next, DGPO takes the encoder outputs H(v) to dynamically construct the purified adjacency matrix A~(v) via differentiable thresholding in Eq. (6).
The updated graph A~(v) is not used to replace A~(v) for forward message passing in subsequent training iterations, which prevents early-stage topological errors from destabilizing GNN propagation. Instead, A~(v) is coupled with the graph encoder during backpropagation through the structural and topology-consistent reconstruction losses. First, because A~(v) is computed directly from the encoder representations H(v) through a differentiable thresholding operator, minimizing the structural entropy loss Lstr backpropagates gradients through A~(v) into H(v) and the encoder parameters, encouraging a compact and cluster-consistent latent manifold where cross-class edges are suppressed. Second, the topology-consistent reconstruction loss Ltopo uses A~(v) to compute random-walk-smoothed features as a clean structural target to supervise the decoded outputs Z^(v), guiding the encoder and decoder to recover masked node attributes from reliable, class-consistent neighbors.