Weaknesses
1. The innovation seems limited, as graph rewiring is already a common approach to tackle heterophilous graphs, as seen in [1] and [2]. However, it would be beneficial for the authors to clarify how their kernel spectral biclustering approach meaningfully differs from existing graph rewiring techniques. Specifically, a comparison of how these methods address heterophily and clustering effectiveness would strengthen the argument for the novelty of their approach. By expanding on the specific advantages or unique contributions of kernel spectral biclustering over rewiring methods, the authors could better situate their work within the current landscape.
2. While the proposed method achieves the best performance in 11 out of 16 cases, the choice of only five baselines may limit the comprehensiveness of this evaluation. I suggest including additional baselines that are well-regarded in heterophilous graph clustering to provide a more robust comparison, like [4][5]. Additionally, reporting on the statistical significance of the observed improvements would help clarify whether these performance gains are practically meaningful or consistent across datasets.
3. More detailed analysis of Table 4 would improve the discussion of the results. Specifically, it would be helpful for the authors to discuss the relative impact of the different components (e.g., the kernel spectral biclustering loss vs. reconstruction losses) in achieving the overall performance. Additionally, an investigation into any observed trends or dependencies among these components would provide insights into the roles they play in model performance.
4. I recommend adding larger datasets, such as Ogbn-arxiv [3], to further validate the proposed method. Testing on larger datasets could help assess the scalability and robustness of the approach and reveal any computational challenges that might arise when handling high-dimensional data. This addition could provide a more comprehensive evaluation and help demonstrate the method's potential for broader applications.
[1] Jiang, W., Gao, X., Xu, G., Chen, T., & Yin, H. (2024, May). Challenging Low Homophily in Social Recommendation. In Proceedings of the ACM on Web Conference 2024 (pp. 3476-3484).
[2] Zheng, Y., Zhang, H., Lee, V., Zheng, Y., Wang, X., & Pan, S. (2023, July). Finding the missing-half: Graph complementary learning for homophily-prone and heterophily-prone graphs. In International Conference on Machine Learning (pp. 42492-42505). PMLR.
[3] Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., ... & Leskovec, J. (2020). Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33, 22118-22133.
[4] Li W Z, Wang C D, Xiong H, et al. Homogcl: Rethinking homophily in graph contrastive learning[C]//Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2023: 1341-1352.
[5] Gu M, Yang G, Zhou S, et al. Homophily-enhanced structure learning for graph clustering[C]//Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2023: 577-586.