Tailoring Self-Attention for Graph via Rooted Subtrees

Attention mechanisms have made significant strides in graph learning, yet they still exhibit notable limitations: local attention faces challenges in capturing long-range information due to the inherent problems of the message-passing scheme, while global attention cannot reflect the hierarchical neighborhood structure and fails to capture fine-grained local information. In this paper, we propose a novel multi-hop graph attention mechanism, named Subtree Attention (STA), to address the aforementioned issues. STA seamlessly bridges the fully-attentional structure and the rooted subtree, with theoretical proof that STA approximates the global attention under extreme settings. By allowing direct computation of attention weights among multi-hop neighbors, STA mitigates the inherent problems in existing graph attention mechanisms. Further we devise an efficient form for STA by employing kernelized softmax, which yields a linear time complexity. Our resulting GNN architecture, the STAGNN, presents a simple yet performant STA-based graph neural network leveraging a hop-aware attention strategy. Comprehensive evaluations on ten node classification datasets demonstrate that STA-based models outperform existing graph transformers and mainstream GNNs. The code is available at https://github.com/LUMIA-Group/SubTree-Attention.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ds4S6/10 · confidence 3/52023-07-05

Summary

The paper introduces Subtree Attention (STA), a new graph attention mechanism that overcomes limitations of existing mechanisms in graph learning. STA combines fully-attentional structure with rooted subtrees, approximating masked global attention under extreme settings. By computing attention weights among multi-hop neighbors directly, STA addresses previous mechanism problems. The authors present STAGNN, a graph neural network that utilizes STA and a hop-aware attention strategy. Extensive evaluations on ten node classification datasets show that STA-based models outperform existing graph transformers and mainstream GNNs.

Strengths

1, This paper proposes a novel multi-hop graph attention mechanism, Subtree Attention (STA), which bridges the fully-attentional structure and the rooted subtree. 2, By incorporating kernelized softmax, they develop an optimized approach for STA, resulting in a linear time complexity. 3, Experiments have shown that STAGNN achieves good performance on Node classification tasks.

Weaknesses

The performance improvement of NAGphormer is limited, as demonstrated in Table 1. For instance, when evaluating photos, NAGphormer achieved a score of 95.49±0.11, while STAGNN achieved 95.64±0.27. The difference between 95.49 and 95.64 is only 0.15, which is smaller than 0.27. A similar trend is observed in the case of CS. Furthermore, in Physics, STAGNN performs worse than NAGphormer. Additionally, it is worth noting that the variance of STAGNN appears to be much larger than that of NAGphormer.

Questions

Why do you only evaluate STA on the node classification task? What about Graph Classification?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

N/A

Reviewer K8yB6/10 · confidence 3/52023-07-06

Summary

This paper introduces a novel multi-hop graph attention mechanism called SubTree Attention (STA) to address the limitations of both local and global attention in Graph Neural Networks (GNNs). STA allows the root node to attend directly to further neighbors in the subtree, enabling it to gather information from the entire rooted subtree within one layer. This mechanism avoids issues associated with deep architectures using local attention layers and captures the hierarchical neighborhood structure more effectively than global attention. The authors propose an efficient algorithm based on kernelized softmax to calculate attention among multi-hop neighbors. They also provide a theoretical analysis of STA, demonstrating its convergence to a degree-masked global self-attention. The paper introduces STAGNN, a multi-hop graph attention network that integrates the STA module into decoupled GCNs, achieving superior performance compared to existing GNNs and graph transformers. Experimental evaluations on node classification datasets confirm the effectiveness of STA and the competitive performance of STAGNN.

Strengths

* The authors propose a strategy to incorporate attention in multi-hop propagation GNNs and get good performances. * Good theoretical analysis is provided.

Weaknesses

* Lack of baselines (BernNet,...). The authors ignore some important baselines of propagation-based GNNs, including BernNet and successive work. * The kernalized attention is a method from Transformer variants. Need citation. * Datasets are not large, should include OGB-datasets, at least ogbn-arxiv

Questions

* How is Graph Transformer applied to node-classification? Is it the pipeline described? Because graph for node-classification is larger than graphs for graph-level tasks? Maybe subgraph sampling is needed? * Two questions about novelty. The STAGNN is an attention-based method to combine multi-hop information of graphs? The difference of using vanilla transformer and STA is that STA has kernalized attention? If not, can you briefly describe the novelty here? * Is the multi-head attention applied in STAGNN? * If the attention will converge to degree-based global attention, why will STA work? Is the convergence achieved in training?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

yes

Reviewer smZ87/10 · confidence 3/52023-07-06

Summary

The present manuscript proposes a novel graph attention layer, that lies in the middle of the local aggregation scheme of message passing and the global (non-structured) nature of a full-attention. The newly proposed Subtree Attention constructs for each node the similarity pairs of key and query matrices from its k-hop neighborhood. Given the transformed matrices, they perform an efficient aggregation throughout the neighborhood hops by considering the random walk adjacency matrix as the transition matrix that is used in the graph attention. They show that given the utilization of the random walk matrix combined with a kernelized softmax function, they reduce the computational cost to $\mathcal{O}(\mathcal{E})$, where $\mathcal{E}$ the number of graph edges. The authors prove theoretically that given a set of assumptions, the Subtree attention can approximate the global attention, showing that the proposed layer is able to tackle over-smoothing and over-squashing problems that come with MPNNs, while bridging the local with the global attention. Their experimental study shows that STAGNN (their method) has a superior performance with respect to the state-of-the-art graph transformer models.

Strengths

- The Subtree attention is a novel and interesting idea on bridging the notions of the local neighborhood attention with the global graph attention. - The theoretical study shows that subtree attention can approximate global attention, which can be very useful for treating over-smoothing in MPNNs. - Empirically, the model STAGNN seems to outperform its competitor baselines. - The paper is well-written and easy to follow.

Weaknesses

- The experimental study is limited to six benchmark datasets, which it is unclear whether long-range interactions play any role, or instead shallower architectures are, also, effective. - It would be very useful testing STAGNN, also, in other datasets (e.g. from OGB), that reportedly depend on long-range interactions, or in synthetic datasets, so that the impact of the proposed model can be shown better.

Questions

- Can the authors show any empirical evidence that the examined datasets can show a need for deeper models? In other words, do these datasets consist of any long-range interactions, or do they consist only of shallow interactions?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors make a discussion on the supplementary material over possible extensions and ways to improve the STAGNN by incorporating it to more models.

Reviewer sk7A5/10 · confidence 3/52023-07-06

Summary

In this paper, the authors propose a novel Graph Transformer called STA-GNN (SubTree-attention-GNN) to address the over-smoothing and over-squashing in the message-passing scheme. Different from the previous Graph Transformer, STA-GNN's attention mechanism is called SubTree Attention (STA), which computes a node's attention according to different levels of its subtree. To overcome the exponentially growing computation cost STA brings, the authors leverage the linearized attention technique raised by [19]. Experiments show that STAGNN outperforms SOTA GNN on node classification tasks.

Strengths

1. The authors provide a thorough demonstration of STA's workflow with a theoretical analysis of how STA addresses over-smoothing and over-squishing. Besides. The authors propose an effective computing algorithm, which decreases STA time complexity. 2. The experiment result outperforms SOTA GCN and Graph Transformer. 3. Ablation studies show that STA is beneficial to global attention in graph learning.

Weaknesses

1. STA's timing performance is unclear. The authors leverage kernelized Softmax to design an efficient algorithm with time complexity O(epsilon) (the number of edges). However, it is hard to compare STA with Graph transformers' (O(N) where N is the number of nodes) from the time complexity aspect, since the number of nodes and edges are independent in an undirected graph. This paper lacks experiments on STA's and Graph Transformers' timing performance. Without evidence, people will worry about if STA can achieve superior performance within a reasonable time limit. 2. Analysis of STA's space complexity is missing. STA efficient algorithm reduces the number of calculations by storing and re-using previous calculation results for future calculation. But there is no discussion on the potential problem of STA's memory cost. A memory cost comparison can better show the efficiency of STA. 3. Hop2Token used in NAGphormer [1] also learns the graph at the level of hops like STA. The experiment results against with NAGphormer does not seem very superior, especially when the experiment is not comprehensive to include a comparison of time and memory cost. 4. Equation (5) doesn't match the context in Figure 1 (a). N is defined as the number of nodes at the start of Section 2. However, figure 1 (a) says the STA_{k} should only attend the k-th hop neighbors, which is conflict with Equation (5). 5. Lack of explanation when introducing the rewriting from equation (3) to equation (4) (in Sec. 2.2). It is not obvious how to rewrite a similarity function with a selected kernel and feature map. [2] is about how to rewrite the similarity function using a linear kernel, which should be cited here. [1] J. Chen, K. Gao, G. Li, and K. He. Nagphormer: Neighborhood aggregation graph transformer for node classification in large graphs. CoRR, abs/2206.04910, 2022. [2] A. Katharopoulos, A. Vyas, N. Pappas, and F. Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 5156–5165. PMLR, 2020.

Questions

1. In section 5.2, the authors claim "In contrast to MP-GNNs, STAGNN maintains robust performance even when the height of the subtree reaches 100". Where is the source of "the performance of MP-GNNs doesn't maintain robust when the height of the subtree reaches 100"?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

No limitations are mentioned in this paper.

Reviewer K8yB2023-08-14

I appreciate the authors' response. Most of my concerns are addressed. The authors need to fix their final version carefully. I will change my score to weak accept.

Authorsrebuttal2023-08-14

Many thanks

Many thanks for your positive remarks. We are committed to further refining our work and making the necessary improvements to address any concerns. Thank you for the opportunity to enhance the quality of our research.

Reviewer sk7A2023-08-19

Thanks to the authors for the response to my concerns. My major concerns are well addressed. I'm willing to increase my score to 5.

Authorsrebuttal2023-08-20

Thank you for your positive feedback

We appreciate the reviewer's positive feedback. We remain dedicated to ongoing improvement. And we thank the reviewer for helping us improve the quality of our paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC