HeteGraph-Mamba: Heterogeneous Graph Learning via Selective State Space Model

1. The problem statement is unclear. The introduction does not effectively clarify whether the focus is on learning heterogeneous graph representations or capturing long-range dependencies. I suggest the authors clearly delineate the specific focus and differentiate between these two aspects in the introduction.

2. Lack of sufficient motivation. The paper introduces challenges related to selective state space models (SSSMs) but fails to explain the necessity of integrating SSSMs. I recommend the authors elaborate on the limitations of existing approaches and provide examples where SSSMs would be particularly beneficial to enhance the motivation.

3. The experimental results are not convincing:

(1) The paper mentions two recent baselines—LMSPS (Li et al., 2024a) and Graph-Mamba (Wang et al., 2024)—but LMSPS is not included in the experiments. The authors should explain why LMSPS was omitted and discuss its relevance for comparison.

(2) Graph-Mamba focuses on long-range graph sequence modeling rather than heterogeneous graph representation learning. The authors should justify the appropriateness of using it as a baseline.

(3) The paper claims that HGMN performs well on large heterogeneous graphs, yet the four datasets used are all standard datasets, not large-scale ones. I suggest the authors consider using specific large-scale heterogeneous graph datasets to better support their claims regarding HGMN's performance.

W1. The motivation for using Mamba on graphs is unclear. Mamba is faster than Transformer only when modeling sequences longer than 2k. W2. Although the paper demonstrates the superiority of the experimental results, it lacks sufficient theoretical analysis of the proposed method, especially in the theoretical basis for improving the efficiency and accuracy of SSSMs. It is recommended to add more mathematical analysis or complexity proof. W3. The previous experiments only verified the effect of HGMN in the node classification task. The applicability and performance of other graph tasks (such as edge prediction and graph classification) have not been tested, which affects its wide applicability. W4. The model hyperparameters are not explained enough, and the analysis of the functions of each module is also lacking. It is recommended to increase the analysis of hyperparameters and add ablation experiments.

1. The motivation is too weak. Since Mamba claims to handle long-range context better and operates in linear time complexity, the paper applies Mamba to heterogeneous graph learning tasks, expecting superior results. This point lacks persuasiveness.
2. Novelty is very limited. The paper shares many similarities with Graph Mamba（KDD 2024）.

2.1 For instance, in the Tokenization section, both fundamentally encode subgraphs to represent nodes, with the difference being that Graph Mamba uses a random walk approach to sample subgraphs, while this paper uses metapath instances.

2.2Additionally, the language style in the Tokenization section of the article is very similar to that of Graph Mamba.

2.3 In this paper, I think the only difference between it and Graph Mamba lies in the inner and outer ordering methods. However, these ordering methods lack persuasiveness. *Why does the inner ordering rely on the number of metapaths as its basis? *Why can’t the number of node degree be used as a basis during inner ordering? *Shouldn’t the importance of different metapaths be considered?

3. Some important detail is missing. For example, in the Metapath Instance Aggregation section, the implementation of the instance encoder is not explained. How is beta implemented in Equation 9?

4. The experiments are not enough. The ablation experiments are not analyzed and are not comprehensive, for example, there is a lack of ablation on the TOKENIZATION and HETEROGENEITY ALIGNMENT sections. There is also a lack of analysis on key parameters in the method.

Reference:*Graph Mamba: Behrouz, Ali, and Farnoosh Hashemi. "Graph mamba: Towards learning on graphs with state space models." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024.

1. The related work lacks targeting the problem that the paper wants to solve. i.e., if you claim the proposed model can do long-range dependency, so your related work should discuss the work that is related to this challenge.
2. The paper has typos and doesn't provide explanations for some symbols, i.e., $\mathcal{N}_s^M$, and spacial at line 172.
3. Additionally, the experiments are unable to be reproducible. it is unclear how metapaths are built. Since the improvement is marginal, it would be better to show the variance of the performance. This paper lacks an ablation study for the components in the model. sin
4. My major concern is that the local information of a node might gradually disappear during the inner and outer updates since the order of these two processes is only determined by the degree.