- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes Band-Masked Spectral Autoencoding (BMSAE), a self-supervised graph representation learning framework that masks and reconstructs graph spectral bands rather than node features or edges. Specifically, BMSAE decomposes node features into multiple spectral bands using a fixed Meyer-wavelet Parseval frame, randomly masks selected bands, and predicts them from the remaining bands through an encoder-decoder architecture. The paper also establishes a reconstruction error bound based on the Parseval property. Experiments demonstrate competitive performance on graph classification and heterophilic node classification benchmarks.
Critical strengths
- The proposed spectral-band masked reconstruction provides an interesting alternative to conventional spatial-domain masking (e.g., GraphMAE) and spectral contrastive learning.
- The paper establishes a reconstruction bound showing that masked-band prediction errors upper-bound the reconstruction error in the original feature space.
- BMSAE achieves the best reported mean performance on five of six graph classification datasets and four of five heterophilic node classification benchmarks.
Critical weaknesses
- Table 3 compares BMSAE against frozen spectral features, but this does not disentangle the contributions of masked-band reconstruction and the nonlinear encoder. An architecture-matched baseline using the same spectral inputs but a different self-supervised objective would better establish the effectiveness of the proposed masking strategy.
- The main experiments use uniform single-band masking, and four of six graph classification datasets select V=2, reducing training to reciprocal prediction between two broad frequency bands. The paper lacks a systematic investigation of how masking multiple bands, varying masking ratios, or selecting different frequency bands affects representation quality.
- The node classification experiments focus exclusively on five heterophilic benchmarks. The paper does not report these results or provide a detailed analysis of when and why spectral-band masking is beneficial.
Questions for authors
Please refer to weaknesses above.
Response to Reviewers and New Ablations
Thank you for taking the time to review our paper and for your feedback.
Reviewer sAiX
We agree that WL is a strong graph-classification baseline. In Table 1, BMSAE improves over WL by 1.4 points on COLLAB, 1.8 on PROTEINS, 2.2 on D&D, and 2.3 on NCI1. Against the strongest reported comparator on each dataset, the differences are +0.6, +1.8, +1.0, and −2.0 points, respectively. We view these as modest gains on three datasets and a clear loss to GraphCL on NCI1.
We chose Meyer filters because they give a fixed set of smooth, overlapping frequency bands that can be evaluated with Chebyshev polynomials. The ideal Parseval property gives exact synthesis and a unit-constant bound from masked-band prediction error to feature reconstruction error. The argument also applies to other Parseval spectral banks. For example, overlapping cosine filters can be normalized so that their squared responses sum to one across the Laplacian spectrum. Whether that choice gives better downstream performance is an empirical question.
The new controls in Appendix C.2 hold the Meyer bank fixed and vary the training objective. BMSAE has a higher mean than both unmasked controls in all eight distinct graph configurations, supporting the contribution of target withholding with the chosen bank.
Reviewer WZEq
We agree that the bound follows from the non-expansiveness of Parseval synthesis. Its role is to connect the pretext loss to feature reconstruction fidelity. We evaluate downstream representation quality empirically with frozen probes.
To address the attribution gap in Table 3, we added both requested controls in Appendix C.2 and linked them from §6.4. The target-visible control uses the same sampled target and single-band reconstruction loss as BMSAE, with every band visible to the encoder. The full-band control also exposes every band and averages reconstruction losses across all bands on each update. Both reuse the selected filter bank, encoder and decoder architectures, training budget, and frozen-probe protocol.
BMSAE has a higher mean than both controls in all eight distinct graph configurations. Across ten node configurations, it has a higher mean than target-visible in nine, with one tie at displayed precision, and a higher mean than full-band in eight. On Roman-empire Linear, the gains are +2.7 accuracy points over target-visible and +1.9 over full-band. The two full-band exceptions are on Minesweeper, where BMSAE trails by 0.4 and 0.2 ROC-AUC points. These results support a contribution from target withholding in most of the tested configurations.
Reviewer qPxs
Contribution of masking. Appendix C.2 adds two architecture-matched controls using the same spectral filter bank and frozen-probe protocol. One exposes all bands and reconstructs the same sampled target as BMSAE. The other exposes all bands and reconstructs every band on each update. BMSAE has a higher mean than both controls in all eight distinct graph configurations. Across ten node configurations, it has a higher mean than target-visible in nine, with one tie at displayed precision, and a higher mean than full-band in eight. Full-band leads on both Minesweeper configurations.
Masking ratio and band choice. We agree that the four V=2 graph configurations reduce to reciprocal low-/high-band prediction. All four have higher mean accuracy than both controls, so this simple setting can benefit from withholding the target. Eight of the ten node configurations use V>2, and the reported comparison still uses uniform single-band masking. The effect of withholding multiple bands or favoring particular frequencies remains open.
Scope of the node results. Our quantitative node-classification evidence covers five heterophilic benchmarks. The benefit varies within that panel. BMSAE is 6.7 accuracy points ahead of GraphMAE2 on Roman-empire and 1.5 points behind it on Amazon-ratings. As noted in §6.3, the homophilic cases we examined showed smaller gains. The current results leave open which graph and feature properties predict when spectral-band masking will help.