[论文精读]Supervised Community Detection with Line Graph Neural Networks-CSDN博客

本文链接：https://blog.csdn.net/Sherlily/article/details/139330314

该博客详细解读了一篇关于使用线图神经网络（LGNN）进行监督社区检测的论文。LGNN通过非回溯操作在边的邻接矩阵上增强图神经网络（GNN），以捕捉无向图中的有向信息流。论文探讨了LGNN的损失函数、实验结果，包括在随机块模型和SNAP真实数据集上的表现，并指出LGNN在处理社区结构和节点分类任务中的潜力。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

论文网址：[1705.08415] Supervised Community Detection with Line Graph Neural Networks (arxiv.org)

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

⭐内涵大量可视化推导

2.5. Line Graph Neural Networks

2.5.1. Graph neural networks using a family of multiscale graph operators

2.5.2. LGNN: GNN on line graphs with the non-backtraking operator

2.5.3. A loss function invariant under label permutation

2.6. Loss landscape of linear GNN optimization

2.7. Experiments

2.7.1. Stochastic block models

2.7.2. Probing the computational-to-statistical threshold in 5-class SBM

2.7.3. Real datasets from SNAP

2.8. Conclusion

3. 知识补充

3.1. Belief propagation

4. Reference List

1. 省流版

1.1. 心得

（1）改论文发表时间较早，许多公式表达没有统一现在的GNN，此文章统一将矩阵表示为大写字母

（2）不过这个有向信息流捕获感觉还有待商榷，还是感觉偏伪信息

2. 论文逐段精读

2.1. Abstract

①They proposed a family of Graph Neural Networks (GNNs)

②Tasks: supervised community detection

③GNN augmentation: non-backtracking operator on the line graph of edge adjacencies

2.2. Introduction

LGNN can capture directed information flow from undirected graphs

2.3. Problem setup

①Task: node classification

②Graph: $G=\left ( V,E \right )$

③Label of nodes: $y:V\to\{1,\ldots,C\}$ , where $C$ denotes the number of communities

④Training set: $\{(G_{t},y_{t})\}_{t\leq T}$

⑤The minimized loss function: $L(\theta)=\frac{1}{T}\sum_{t\leq T}\ell(\Phi_\theta(G_t),y_t)$ where the predicted label is $\hat{y}=\Phi_{\theta}(G)$ , $\ell$ is a loss function, and $\Phi$ represents the model

2.4. Related works

Briefly introducing GNN and its variants

2.5. Line Graph Neural Networks

2.5.1. Graph neural networks using a family of multiscale graph operators

（1）Line graph construction

①Node feature: $x_i \in \mathbb{R}^{1\times b}$ , $X \in \mathbb{R}^{\left | V \right | \times b}$

②Adjacency matrix: $A \in \mathbb{R}^{V \times V}$ with 1 and 0

③Degree matrix: $D_{ii}=diag\sum_{j\in \left | V \right |}^{}A_{ij}$

④Power graph adjacency matrices: $A_{J}=\operatorname*{min}(1,A^{2^{J}})$ where $J\in\mathbb{N}^{*}$ , the power denotes the number of hop

⑤Identity matrix: $I$

⑥Family matrices: $\mathcal{F}=\{I,D,A,A_{2},...,A_{J}\}$

⑦One GNN layer:

$X^{k+1} \in \mathbb{R}^{\left | V \right | \times b_{k+1}}=GNN\left ( X^k \in \mathbb{R}^{\left | V \right | \times b_{k}} \right )$

⑧GNN methods:

$Z^{(k+1)}=\rho\left[\sum_{O_i\in\mathcal{F}}O_iX^{(k)}\theta_i\right],\quad\overline{Z}^{(k+1)}=\sum_{O_i\in\mathcal{F}}O_iX^{(k)}\theta_i$

are two terms in one layer, where $\theta_{j}\in\mathbb{R}^{b_{k}\times\frac{b_{k+1}}{2}}$ denotes learnable parameters, $\rho \left ( \cdot \right )$ there is ReLU, further update $X^k$ by: