Long-range Brain Graph Transformer

Summary

This paper employs a random walk approach to capture long-range dependencies in the brain through a feature engineering scheme. The computed adaptive factors are incorporated into node features and used to train a Transformer model.

Strengths

1. The idea of capturing long-range dependencies in the brain is reasonable. 2. The paper is well-written and easy to understand. 3. The experimental results are strong on the two datasets tested.

Weaknesses

1. Some design choices in the proposed method are not well-motivated: - Using the designed adaptive factors (Eq. 1) instead of Pearson correlations. - Introducing the degree matrix in the random walk kernel. 2. The parameter \( k \) in Section 3.2 is not clearly defined; it is assumed that \( k = K \). 3. The descriptions of the dataset construction and experimental settings are unclear. Specifically: - Which atlas is used, and how many ROIs does it contain? - What exact value is used to threshold the connectivity to obtain the adjacency matrix? - Given that the number of hops is set up to 32, it is expected that the adjacency matrix is super sparse. However, such a high threshold is uncommon in brain network analysis. 4. The paper conducts experiments on only two datasets, which is insufficient. The ADNI dataset used contains only two classes with 130 subjects. It would be better to extend the experiments to include more datasets as referenced in [1], which includes six datasets and over 1.3k subjects for ADNI. 5. There is a lack of comparison with the latest related works. For example: - [2] introduces a graph augmentation using ALFF features. - [3] presents a global matching-based graph kernel that captures dynamic changes in evolving brain networks. - [4] utilizes clustering-based graph pooling for readout. 6. It would be valuable to incorporate the proposed ALGA method with GNN baselines such as GCN and GAT to evaluate whether it can improve their performance. 7. The paper lacks deeper analysis from a neuroscience perspective. It is important to provide insights from the model results, such as identifying specific ROIs related to the target disease and whether these findings align with medical literature. References: [1] Data-driven network neuroscience: On data collection and benchmark. NIPS 2023 [2] A-GCL: Adversarial graph contrastive learning for fMRI analysis to diagnose neurodevelopmental disorders. MIA 2023 [3] Effective Graph Kernels for Evolving Functional Brain Networks. WSDM 2023 [4] Contrastive Graph Pooling for Explainable Classification of Brain Networks. IEEE TMI 2024

Questions

1. Which atlas is used in the dataset? 2. Which ROIs obtain high attention, and do these match domain observations for the diseases studied?

Rating

5

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

The limitations have been discussed in this paper.

Summary

This paper highlights a significant gap in the existing literature on brain network representation learning, specifically the inadequacy of current methods to effectively capture long-range dependencies, leading to limited an integrated understanding of brain-wide communication. To bridge this gap, the paper introduces ALTER, an innovative model designed for adaptive awareness of long-range dependencies and brain network representation learning. The model consists of a long-range aware strategy for capturing long-range dependencies and a transformer framework for integrating multi-level communication understanding.

Strengths

(1) This paper is generally clear and well-written, providing a comprehensive analysis of the problem to be solved, supported by some convincing references, and emphasizing the urgent need to come up with solutions that can capture long-distance dependencies between brain ROIs. At the same time, the authors make the problem better understood with illustrations. (2) The logic of the paper is smooth. The authors clarify the issue, after which they elaborate on the theoretical causes of the issue, and finally design a specific biased random walk strategy for the theoretical causes and obtain encouraging experimental results. (3) I found the method presented here is technically sound with excellent results. From the experimental results, it is shown that the proposed adaptive long-range aware strategy is very effective in long-range dependencies capture and can greatly enhance the effectiveness of the disease diagnosis task. In addition, the ablation study conducted on various components reveals the superiority of the adaptive long-range aware strategy. (4) The authors clearly and frankly understand the drawbacks and strengths of the method, e.g. the method does not allow for a good trade-off between long- and short-range. This makes this method a clear one to build on top of, prompting further interesting work in highly important domain.

Weaknesses

1. Although the results of the ablation experiments verify that biased random walk is efficient, it is difficult to convince me of the necessity of introducing this module based only on the theoretical description of biased random walk in the methods section. As far as I know, random walk is common in graph representation learning. Therefore, I would like to ask the authors to provide me with a more in-depth analysis of biased random walk, otherwise I do not understand why biased random walk is effective in capturing long-range dependencies in brain networks. 2. Transformer-based graph representation learning is not original in brain network analysis. It is recommended that the authors provide a specific explanation for fusing long-range and short-range dependent embeddings using the graph Transformer. 3. Lack of interpretable analysis. Interpretive analysis of brain network representation learning models geared toward brain disease diagnosis is important. In the context of brain network analysis, I suggest that papers should explain and identify the brain regions or networks that are most relevant to the task of brain disease classification.

Questions

1. Have the authors considered tasks other than binary categorization? 2. Am I correctly understanding that the dimension of the original feature embedding and the long-range dependency embedding after splicing is equal to the original feature embedding dimension? In the representation in Figure 2, I get the impression that the dimension after splicing is unchanged, but this does not seem to be the case from the text. Could the authors clarify this point? 3. The authors conducted experiments on fMRI data sets, and fMRI is only single-modal. Is the biased random walk strategy of ALTER still valid on multi-modal data?

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

See weaknesses and questions for details.

Summary

The study employs the adaptive long-range aware graph transformer (ALTER) to tackle the challenge of weak comprehension in whole-brain communication, which arises from the failure to capture long-range dependencies in brain networks. Initially, the study encodes long-range dependencies into long-range embeddings through biased random walk sampling, thereby enriching the embeddings with information on long-range dependencies. Subsequently, the study takes into account the significance of both short- and long-range dependencies in brain network analysis tasks. It introduces the graph transformer, which uses a self-attention mechanism to integrate these dependencies between brain ROIs. The objective is to capture varying levels of communication connections within the human brain. Experimental results demonstrate that ALTER outperforms current SOTA graph learning baselines, achieving superior AUC and ACC scores on the ABIDE and ADNI datasets.

Strengths

(1) This work is significant. How to facilitate the understanding of communication and information processing among brain ROIs is a key issue. The authors provide a clear overview of the necessity of long-range dependencies for understanding communication and information processing among brain ROIs, and propose an effective method to capture long-range dependencies in brain networks. (2) The paper is aesthetically pleasing in its writing form, especially the diagrams and charts that make it easy for the reader to understand the exact process of the work done. Meanwhile, the authors provide detailed preprocessing and implementation details. As a result, this paper is highly reproducible. (3) The paper provides complete experimental results. Besides comparisons with the baseline and ablation studies on modules, the appendix section validates the state-of-the-art of the adaptive long-range dependency aware strategy on multiple readout functions, further enhances reproducibility, and facilitates a deeper understanding of the method. (4) ALTER is an ingeniously crafted framework that is able to adaptively perceive both long- and short-range dependencies in brain networks. In addition, ALTER interestingly takes inter-ROI correlations into account in the capture of long-range dependencies in brain networks.

Weaknesses

I have identified three main weaknesses that need to be clarified during the rebuttal process, which is the reason I gave a weak acceptance despite the strengths. If my points are properly addressed, I will be happy to review my scores based on the results of the rebuttal process. I numbered the comments to facilitate discussion. 1. The technical contributions are neutral. The proposed ALTER seems to be a combination of the random walk and the graph transformer, but its novelty and difficulty are unclear. 2. ALTER obtained SPE scores that were 5% lower than BrainNETTF on the ADNI dataset and SEN scores that were 1.3% lower than Graphormer on the ABIDE dataset. This performance lag is significant, but the authors do not seem to have clearly explained the reasons for the SPE and SEN lags. 3. The authors state in line 202 that ALERT uses a linear layer to inject long-range dependencies into the brain graph transformer, but that direct collocation also accomplishes the above, and the authors do not state the necessity of introducing a linear layer.

Questions

In addition to what I wrote in the "Weaknesses" section, I have a couple of questions: 1. How does the author define long-range dependence and short-range dependence? 2. Did the authors attempt to choose other correlation measures than the Pearson correlation coefficient to define the adaptive factors? 3. Could you elaborate on the need to introduce the graph transformer in the paper? Would replacing the graph transformer with a GNN affect the results? 4. Could the authors elaborate on the correlation between the attention map and the example graph in Figure 5?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper emphasizes the possible negative social impact of work. However, in my opinion, the negative societal impact is not only medical errors due to prediction errors, but should also include the negative ethical impact of the model. Adding ethical concerns would help the reader better understand the potential impact of the proposed methodology.

Summary

This work proposes Adaptive Long-range aware TransformER (ALTER), a brain graph transformer to capture long-range dependencies between brain ROIs utilizing biased random walk.

Strengths

1. This work introduces a novel brain graph transformer with adaptive long-range awareness, which leverages the communication strengths between ROIs to guide the capturing of long-range dependencies. 2. The result demonstrates that ALTER consistently outperforms generalized graph learning methods and other graph learning-based brain network analysis methods.

Weaknesses

1. Even though the proposed ALTER is better than all selected baselines, some baselines like BrainGNN can provide both the prediction and interpretation. It is unclear whether the ALTER can also explain the important disease-specific pattern and find the biomarkers. 2. It is unclear how to implement the comparable baselines and how to build the brain graph for them. 3. It is unclear how to preprocess the fMRI in detail and how to conduct the quality control.

Questions

1. What is the final loss function for this task? 2. How many times do you repeat the experiment? Have you conducted k-fold cross-validation to examine the result? 3. How about the ALTER’s ability to capture long-range dependencies in the patient's group instead of choosing one example to show in Figure 4? Can it get the same conclusion in the group-level analysis?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Follow-up Questions

I am glad to find that the authors have addressed most of my concerns. However, I still have some follow-up questions: 1. Regarding the threshold used for sparsifying the connectivity to obtain the adjacency matrix, does 0.3 mean retaining edges with Pearson correlation larger than 0.3 or keeping the top 30% of edges? Additionally, does this thresholding drop all negative edges, or are absolute values used? 2. For the new experiment conducted, are you using the same settings as the original paper, or have you applied your own settings to them?

Reply to rebuttal

I appreciate the detailed clarifications and additional interpretable experiments provided. These have effectively addressed my previous concerns and underscored the significance of this work. Of course, I am inclined to increase my score and recommend acceptance of the paper. Overall, the manuscript is well written and represents a valuable contribution to the research field. I look forward to seeing the integration of the aforementioned discussion in the next version, particularly the interpretability analysis.