- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper revisits attribute-missing graph clustering through a benchmark covering 22 datasets and 32 baselines. It proposes VISTA, a lightweight framework combining bounded attribute recovery, graph smoothing, and cross-view prototype learning. Extensive experiments demonstrate competitive clustering performance, robustness, and scalability.
Critical strengths
- The paper is generally well written and easy to follow, with a clear overall structure.
- The experiments are comprehensive, covering different datasets, missing rates, missingness mechanisms, ablations, and computational efficiency.
- The supplementary material includes code for VISTA and the baselines, supporting reproducibility and inspection of the experimental settings.
Critical weaknesses
- Figure clarity. Some visual details need attention. In particular, the feedback arrows on the right side of Figure 2 have unclear endpoints and directions. Their meaning should be made explicit.
- Limited methodological explanation. Section 4 presents the main equations but provides insufficient discussion of the design motivations and connections between stages. A step-by-step algorithm or pseudocode would improve clarity.
- Benchmark protocol needs clarification. Section 5.1 uses published configurations or default settings for baselines without specifying comparable tuning budgets and model selection criteria. These details matter for interpreting the central finding that specialized methods do not necessarily outperform conventional alternatives.
Questions for authors
Please see weaknesses.
2 comments
Response
Thanks for your comment, I maintain my positive score.
We sincerely thank the reviewer for the positive assessment and the concrete suggestions.
W1. Clarity of Figure 2
Figure 2 now shows every step of the feedback loop explicitly. We redrew the iterative-feedback panel of Figure 2. Every arrow now connects two stages directly, and a labeled Feedback arrow from Refinement back to Imputation closes the imputation–clustering loop. A new sentence in the caption, marked in blue, states what the loop means.
The framework figure now presents VISTA as one connected pipeline. We also redesigned the framework figure (Figure 3). It is now organized into four numbered stages, and each stage is split into sub-modules annotated with their equations. Labeled arrows trace the data flow from the observed attributes to the final clustering, and the parameter-free and trainable parts are marked explicitly.
W2. Design motivation in Section 4 and pseudocode
Section 4 now explains how each stage feeds the next and why it is built this way. We added the design motivation of each stage and the connections between stages to Section 4, and Algorithm 1 to the new Appendix K. All additions are marked in blue.
Each stage supplies what the next one needs. Section 4 now opens with an overview of the four stages. Recovery gives missing nodes attribute evidence without a learnable imputer. Smoothing places observed and recovered nodes in a common neighborhood-aggregated space and supplies the topology source of the second view. Cross-view prototype learning turns the two views into clustering-oriented embeddings, and KMeans partitions their fusion.
Each design choice follows from what the stage needs to provide. Recovery writes each missing row once, so the row stays anchored to the nearest observed evidence, and the bound S limits how far recovered evidence travels. H and T contain no learnable parameters, so both are precomputed and training involves no graph propagation. At inference the second view is built without dropout, so the fused embedding combines the base evidence of H with the topology-enriched evidence of T.
Algorithm 1 lists the full procedure, from the normalized adjacency and bounded recovery to training and the final KMeans:
W3. Benchmark protocol: tuning budget and model selection
The central finding compares baseline families under one protocol, so it does not depend on how any single family is configured. All baselines run their official code with the per-dataset configurations released by their authors, or with default settings when none are released. Iterative baselines use the checkpoint rule of their official code, and all methods share the same missing masks and seeds 41-45. No family receives a larger tuning budget than another. We added this protocol to Section 5.1 of the revised paper.
No baseline is reported in a failed state. When a baseline collapsed on a dataset that its original paper does not cover, we gave it an extended tuning budget and reported its best configuration instead of the collapsed run. This step can only favor the baselines, and the specialized methods still do not outperform the conventional ones.
The finding also holds without model selection. We evaluated every iterative method at the final epoch of its training schedule, with no label-based selection (new Appendix I). The results in Table 1 show that the best attribute-missing method remains below the best conventional and scalable methods, as in the main results.
Table 1. Average final-epoch ACC (%) over the 17 datasets, two representative methods per family.
We thank the reviewer again and would be glad to clarify anything further.