- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes SAMA, a graph contrastive learning framework that jointly considers graph structure and node attributes during data augmentation. The method combines deep non-negative matrix factorization (DNMF) and spectral feature filtering to construct multiple augmented graph views, followed by contrastive representation learning. Experiments on six benchmark datasets demonstrate improvements in node classification over sixteen baseline methods.
Critical strengths
S1. Comprehensive experimental comparisons. The authors evaluate SAMA on six widely used graph datasets and compare it against sixteen baselines, including several representative graph contrastive learning methods. The reported results consistently outperform the compared approaches.
S2. Component-level analysis. The paper provides ablation experiments investigating the contributions of structural information, attribute information, and regularization. The results demonstrate that combining these components improves downstream performance.
Critical weaknesses
W1. Limited methodological novelty compared with existing graph contrastive learning approaches.
The proposed framework primarily combines existing techniques, including DNMF, spectral feature filtering, and InfoNCE-based contrastive learning. The idea of jointly preserving structural and attribute information has already been investigated in previous studies, such as ASP (AAAI 2023), which is also included as a baseline. The paper does not sufficiently explain the fundamental methodological differences between SAMA and these closely related approaches. Consequently, the novelty appears to lie mainly in the integration of existing components rather than a new learning principle.
W2. The theoretical guarantee of semantic preservation is insufficiently justified.
Theorem 1 claims that DNMF guarantees similar latent representations for nodes within the same semantic cluster while separating nodes from different clusters. However, the proof in Appendix C relies on strong block-diagonal assumptions and informal arguments. In particular, the proof assumes that nodes within the same cluster have identical connectivity patterns, which does not generally follow from a block-diagonal adjacency matrix. Moreover, the extension from single-layer NMF to DNMF is asserted without a rigorous derivation. The claimed perturbation stability also lacks explicit conditions and bounds. These issues undermine the validity of the paper's central theoretical contribution.
W3. Potential concerns about label information and experimental fairness.
Section 4.1 specifies that the intermediate representation dimension (r) is set to the number of classes, while the method is evaluated as an unsupervised representation learning approach. Although using the number of classes does not necessarily constitute label leakage, this setting requires clarification regarding whether ground-truth label information is used during model configuration. The authors should explain how (r) is selected when class information is unavailable and whether the same degree of dataset-specific tuning is permitted for competing methods. Sensitivity analysis with label-independent choices of (r) is also necessary.
W4. Incomplete computational complexity analysis and scalability evaluation.
The complexity analysis in Appendix C assumes efficient spectral filtering and negative sampling, but the implementation details do not establish whether these approximations are actually used. In particular, the contrastive objective in Eq. (8) considers all other nodes as negatives, potentially requiring quadratic pairwise similarity computations. Furthermore, the reconstructed adjacency matrices may become dense, substantially increasing memory and computation requirements. The paper provides no empirical runtime, memory consumption, or scalability comparisons with existing methods.
W5. Insufficient evidence that the generated augmentations preserve meaningful graph semantics.
Although semantic preservation is the central motivation, the proposed consistency objective in Eq. (6) measures reconstruction differences using Frobenius norms. Such numerical similarity does not necessarily guarantee preservation of meaningful graph semantics. The experiments focus primarily on downstream node classification accuracy, without directly evaluating semantic consistency, structural preservation, or augmentation diversity. Therefore, it remains unclear whether the improvements arise from semantic preservation or simply from additional structural smoothing and regularization.
Questions for authors
Q1. What are the fundamental methodological differences between SAMA and ASP (AAAI 2023), particularly regarding joint structure-attribute modeling and semantic preservation?
Q2. Can the authors provide a rigorous proof of Theorem 1 under explicitly stated assumptions? In particular, how does the proof extend from single-layer NMF to arbitrary-depth DNMF?
Q3. Why is the intermediate dimension (r) set to the number of classes in an unsupervised learning setting? How does performance change when (r) is selected independently of ground-truth class information?
Q4. Does the implementation explicitly construct dense augmented adjacency matrices and compute all-pairs contrastive similarities? Please report actual runtime and peak GPU memory usage.
Q5. Can the authors directly quantify semantic preservation and augmentation diversity, rather than relying exclusively on downstream classification accuracy?