您的位置:首页 > 游戏 > 手游 > [论文精读]Polarized message-passing in graph neural networks

[论文精读]Polarized message-passing in graph neural networks

2024/12/23 8:10:50 来源:https://blog.csdn.net/Sherlily/article/details/142237103  浏览:    关键词:[论文精读]Polarized message-passing in graph neural networks

论文网址:Polarized message-passing in graph neural networks - ScienceDirect

论文代码:he-tiantian/PMP-GNNs:极化消息传递图神经网络的 Pytorch 实现,发表在 Artificial Intelligence,2024 年。 (github.com)

英文是纯手打的!论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误,若有发现欢迎评论指正!文章偏向于笔记,谨慎食用

目录

1. 省流版

1.1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. The polarized message-passing paradigm

2.3.1. Notations

2.3.2. Formulating polarized message-passing at a single layer

2.3.3. Expressiveness of polarized message-passing

2.4. Polarized message-passing graph neural networks

2.4.1. Polarized message-passing graph convolutional network

2.4.2. Polarized message-passing graph attention network

2.4.3. Polarized message-passing graph PageRank network

2.4.4. Loss function

2.4.5. Computational complexity of PMP-GNNs

2.4.6. Expressiveness of PMP-GNNs

2.5. Theoretical analysis

2.5.1. Proof of Theorem 1

2.5.2. Prood of Theorem 2

2.6. Related work

2.6.1. Graph neural networks

2.6.2. Message-passing graph neural networks

2.7. Experimental setup

2.7.1. Comparison baselines

2.7.2. Datasets and real-world analytical tasks

2.7.3. Experimental settings

2.8. Results and analysis

2.8.1. Classification performance on scientific articles

2.8.2. Clustering performance on blogs and text-based images

2.8.3. Analytical performance in coauthor graphs

2.8.4. Performance in rich-text classification

2.8.5. Comparisons of similarity-, dissimilarity-, and PMP-based GNNs

2.8.6. Comparisons of learned message weights

2.9. Conclusion

3. Reference


1. 省流版

1.1. 心得

(1)好多数学证明,看起来挺严谨的样子。为什么这样的东西只值二区,我感觉挺好的。可能是创新不是特别大?但是实验属于是多麻了

2. 论文逐段精读

2.1. Abstract

        ①They proposed Polarized message-passing (PMP) to capture the similarity relationship between nodes

        ②Thus, three new GNNs were derived, PMP graph convolutional network (PMP-GCN), PMP graph attention network (PMP-GAT), and PMP graph PageRank network (PMP-GPN)

        ③They verified them on 12 real datasets based 5 tasks

2.2. Introduction

        ①Current message-passing graph neural networks (MPGNNs) consider the SIMILARITY more, thus the authors want to contain the DISSIMILARITY

        ②⭐They decomposed traditional adjacency matrix A to 2 matrices, the SIMILARITY matrix and the DISSIMILARITY matrix

        ③⭐They combine the two sim matrices to new sparse matrix, which is the normalized product of the two sim matrices(咦~感觉这样不能真正实现复合?反而揉在一起了,怎么体现极化呢

negate  vt.否定;否认;使无效;取消

2.3. The polarized message-passing paradigm

2.3.1. Notations

        ①Graph is represented as G=\{V,E,\mathbf{X}\}, where each graph contains N nodes and \left | E \right | edges

        ②\mathbf{X}\in\mathbb{R}^{N\times D} denotes the input node feature matrix, \mathbf{H}^{l} is the node feature matrix of the subsequent convolved, \mathbf{H}^{1}=\mathbf{X}

        ③each node belongs to a class of C

        ④\mathcal{N}_{i} denotes one-hop neighbor 

        ⑤\mathbf{A}\in\{0,1\}^{N\times N} denotes the adjacency matrix

        ⑥\mathbf{W}^{l} denotes the learnable weight matrix

        ⑦\mathbf{p}^{l} denotes the latent positions at the l-th layer

2.3.2. Formulating polarized message-passing at a single layer

        ①Node correlations matrix \mathbf{C}:

\begin{aligned}&\mathbf{H}=\mathbf{H}^{l}\mathbf{W}_{h}^{l},\mathbf{P}=\mathbf{P}^{l}\mathbf{W}_{p}^{l},\\&\mathbf{C}_{ij}=\mathbf{A}_{ij}\cdot[\mathbf{a}(\mathbf{H}_{i,:}||\mathbf{H}_{j,:})^{T}+\mathbf{P}_{i,:}diag(\mathbf{b})\mathbf{P}_{j,:}^{T}]\end{aligned}

where the subscript _{i, :} denotes the i-th row vector, || is the concatenation symbol, \mathbf{a} and \mathbf{b} are the learnable weight vectors. (C= A_{ij}*{[a,1]*[1,2f] + [N,p]*[p,p]*[p,N]},感觉上这里A是一个二值化数,难道这里会强行让f=1/2N吗)

        ②Node uncorrelated matrix \mathbf{S}:

\mathbf{S}_{ij}=\mathbf{A}_{ij}\cdot[\alpha|\mathbf{H}\mathbf{W}_{i,:}-(\mathbf{H}\mathbf{W})_{j,:}|^2+(1-\alpha)|\mathbf{P}_{i,:}-\mathbf{P}_{j,:}|^2]

where \alpha\in(0,1) is the learnable parameter

        ③Construction of integrated polarization information:

\begin{aligned} &\mathbf{E}_{ij}=\mathbf{A}_{ij}\cdot\exp(\mathbf{C}_{ij}-\beta\mathbf{S}_{ij}), \\ &\mathbf{M}_{i}=\sum_{j\in\mathcal{N}_{i}}\phi(\mathbf{E}_{ij})\cdot\mathbf{H}_{j,:}, \\ &\mathrm{FEATURE~UPDATE:} \mathbf{H}_{i,:}^{l+1}=\frac{\theta}{|\mathcal{N}_{i}|}\mathbf{H}_{i,:}+\mathbf{M}_{i}, \end{aligned}

where \phi \left ( \cdot \right ) is an appropriately function leveraging \mathbf{E} to generate the weights for computing messages from neighbors, \beta denotes learnable positive parameter, \theta denotes learnable positive irrational

        ④Polarized message-passing (PMP):

2.3.3. Expressiveness of polarized message-passing

        ①大概证明了一下为什么他们提出来的极化消息传递更好,如有需要可以参考原文

2.4. Polarized message-passing graph neural networks

2.4.1. Polarized message-passing graph convolutional network

        ①PMP-GCN:

\begin{aligned} &\mathbf{H}=\mathbf{H}^l\mathbf{W}_h^l,\mathbf{P}=\mathbf{P}^l\mathbf{W}_p^l,\mathbf{E}_{ij}=\mathbf{A}_{ij}\cdot\exp(\mathbf{C}_{ij}-\beta\mathbf{S}_{ij}), \\ &\mathbf{M}_{i}={\frac{1}{\sqrt{\sum_{k\in\mathcal{N}_{i}}\mathbf{E}_{ik}}}}\sum_{j\in\mathcal{N}_{i}}{\frac{\mathbf{E}_{ij}}{\sqrt{\sum_{k\in\mathcal{N}_{j}}\mathbf{E}_{jk}}}}\cdot\mathbf{H}_{j,:}, \\ &\mathbf{H}_{i,}^{l+1}={\frac{\theta}{|{\mathcal{N}}_{i}|}}\mathbf{H}_{i,:}+\mathbf{M}_{i}. \end{aligned}

2.4.2. Polarized message-passing graph attention network

        ①PMP-GAT:

\begin{aligned} &\mathbf{H}=\mathbf{H}^{l}\mathbf{W}_{h}^{l},\mathbf{P}=\mathbf{P}^{l}\mathbf{W}_{p}^{l},\mathbf{E}_{ij}=\mathbf{A}_{ij}\cdot\exp(\mathbf{C}_{ij}-\beta\mathbf{S}_{ij}), \\ &\mathbf{M}_{i}=\sum_{j\in\mathcal{N}_{i}}\frac{\mathbf{E}_{ij}}{\sum_{k\in\mathcal{N}_{i}}\mathbf{E}_{ik}}\cdot\mathbf{H}_{j,:}, \\ &\mathbf{H}_{i,}^{l+1}=\frac{\theta}{|\mathcal{N}_{i}|}\mathbf{H}_{i,:}+\mathbf{M}_{i}. \end{aligned}

2.4.3. Polarized message-passing graph PageRank network

        ①PMP-GPN:

\begin{aligned}&\mathbf{H}^{in}=MLP(\mathbf{X}),\mathbf{P}^{in}=MLP(\mathbf{P}^{0}),\mathbf{E}_{ij}=\mathbf{A}_{ij}\cdot\exp(\mathbf{C}_{ij}-\beta\mathbf{S}_{ij}),\\&\mathbf{M}_{i}=\frac{1}{\sqrt{\sum_{k\in\mathcal{N}_{i}}\mathbf{E}_{ik}}}\sum_{j\in\mathcal{N}_{i}}\frac{\mathbf{E}_{ij}}{\sqrt{\sum_{k\in\mathcal{N}_{j}}\mathbf{E}_{jk}}}\cdot\mathbf{H}_{j,,},\\&\mathbf{H}_{i,}^{l+1}=\frac{\theta}{|\mathcal{N}_{i}|}\mathbf{H}_{i,:}+(1-\alpha)\mathbf{M}_{i}+\alpha\mathbf{H}_{i,:}^{in},\end{aligned}

2.4.4. Loss function

        ①Loss function contains task-specific loss and latent-position loss:

L=L_{task}+tr(\mathbf{P}^{outT}\mathbf{LP}^{out})

where \mathbf{L} denotes the Laplacian matrix of \mathbf{A}

2.4.5. Computational complexity of PMP-GNNs

        ①Mapping computational complexity: O(2Ndd^{\prime})

        ②Computing \mathbf{C} and \mathbf{S}O(4|E|d^{\prime})

        ③Representation learning: O(|E|(1+d^{\prime}))

        ④Layer complexity of PMP-GCN and PMP-GAT: O(|E|(1+5d')+2Ndd')

        ⑤Layer complexity of PMP-GPN: O(|E|d)

2.4.6. Expressiveness of PMP-GNNs

        ①作者证明他们的PMP可以通过Weisfeiler-Lehman 检验,并且说所有MPGNN表达性的上限是WL

——————————————以下证明暂略——————————————

2.5. Theoretical analysis

2.5.1. Proof of Theorem 1

2.5.2. Prood of Theorem 2

(1)Schematic methodology to complete the proof of Theorem 2

(2)The equivalence between the PMP-GCN layer and the 1-WL test

(3)The equivalence between the PMP-GAT layer and the 1-WL test

(4)The equivalence between the PMP-GPN layer and the 1-WL test

2.6. Related work

2.6.1. Graph neural networks

        ①GNN can be used to scientific document classification, coauthor analysis, conversation generation, social community detection, and text classification

2.6.2. Message-passing graph neural networks

        ①List some strategies

2.7. Experimental setup

2.7.1. Comparison baselines

        ①Compared model: a) GNN: GCN, MoNer, GraphSAGE, ARMA; b) Attention based GNN: GAT, GATv2, HardGAT, MAGNA; c) Diffused (PageRank based) GNN: JKNet, APPNP, SGC, GPR

2.7.2. Datasets and real-world analytical tasks

        ①Introduced each dataset:

2.7.3. Experimental settings

        ①Specifically list the parameters in compared models

        ②Hidden layer: 35 for PMP-GCN, 64 for PMP-GAT and PMP-GPN

        ③Dropout rate: 0.75 for PMP-GCN, 0.6 for PMP-GAT, 0.5 for PMP-GPN

        ④Learning rate: 0.01 for PMP-GCN and PMP-GPN, 0.005 for PMP-GAT

        ⑤Attention head: only 1 for PMP-GAT

        ⑥Optimizer: Adam

        ⑦Epoch: 10

        ⑧Performance: average

        ⑨Introduced the distribution of train/val/test set

2.8. Results and analysis

2.8.1. Classification performance on scientific articles

        ①Comparison on traditional GNN:

        ②Comparison on attention based GNN:

        ③Comparison on PageRank based GNN:

2.8.2. Clustering performance on blogs and text-based images

        ①Comparison on traditional GNN:

        ②Comparison on attention based GNN:

        ③Comparison on PageRank based GNN:

2.8.3. Analytical performance in coauthor graphs

        ①Comparison on traditional GNN:

        ②Comparison on attention based GNN:

        ③Comparison on PageRank based GNN:

2.8.4. Performance in rich-text classification

        ①Comparison on traditional GNN:

        ②Comparison on attention based GNN:

        ③Comparison on PageRank based GNN:

2.8.5. Comparisons of similarity-, dissimilarity-, and PMP-based GNNs

        ①Ablation studies on 3 types of GNN:

2.8.6. Comparisons of learned message weights

        ①Learned message weights blation and accumulative distribution of normalized neighbor distances regarding node features and graph structure:

2.9. Conclusion

         ....怎么做了两三百个实验啊

3. Reference

He, T. et al. (2024) 'Polarized message-passing in graph neural networks', Artificial Intelligence, 331. doi: Redirecting

版权声明:

本网仅为发布的内容提供存储空间,不对发表、转载的内容提供任何形式的保证。凡本网注明“来源:XXX网络”的作品,均转载自其它媒体,著作权归作者所有,商业转载请联系作者获得授权,非商业转载请注明出处。

我们尊重并感谢每一位作者,均已注明文章来源和作者。如因作品内容、版权或其它问题,请及时与我们联系,联系邮箱:809451989@qq.com,投稿邮箱:809451989@qq.com