论文解读二代GCN《Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering》

Python微信订餐小程序课程视频

https://edu.csdn.net/course/detail/36074

Python实战量化交易理财系统

https://edu.csdn.net/course/detail/35475

Paper Information

Title：Convolutional Neural Networks on Graphs with Fast Localized Spectral FilteringAuthors：Michaël DefferrardXavier BressonPierre VandergheynstPaper：Download Source：NeurIPS 2016

---

Abstract

基于   spectral graph theory  ，为设计 localized convolutional filters on graphs  提供了必要的数学背景和有效的数值方案。

Our method’s advantages：

• • The proposed technique offers  linear computational complexity.
    • Constant learning complexity.

---

Review

At the begining ,we just do a review about the first generation GCN.

xk+1,j=h(V∑i=1fk−1Fk,i,jVTxk,i)(j=1…fk)xk+1,j=h(V∑i=1fk−1Fk,i,jVTxk,i)(j=1…fk)x_{k+1, j}=h\left(V \sum \limits _{i=1}^{f_{k-1}} F_{k, i, j} V^{T} x_{k, i}\right) \quad\left(j=1 \ldots f_{k}\right)

其中，F(k,i,j)F(k,i,j)F(k,i,j) 是图卷积神经网络在第  kkk  层的 filter ，可学习，具有任意性。由图卷积公式可以看出，卷积核不具备局部性，并且依赖于图的 Laplacian 的谱分解。

---

1 Introduction

CNN的优势：

• • Offer an efficient architecture to extract highly meaningful statistical patterns in large-scale and high-dimensional datasets.
    • Can learn  local stationary structures.
    • Can  compose the local stationary structures to form multi-scale hierarchical patterns in many domain.

What CNN do?

• • CNNs extract the local stationarity property of the input data or signals by revealing local features that are shared across the data domain.
    • Similar features are identifified with localized convolutional fifilters or kernels,which are learned from the data.

Why CNN can recognize identical features independently of their spatial locations?

• • Due to  Convolutional filters are shift  or translation-invariant filters.

Then,explain the common of  non-Euclidean data ,such as social networks, gene data on biological regulatory networks, log data on telecommunication networks, or text documents on word embeddings

Why CNN can’t apply in graph directly?

• • The convolution and pooling operators are only defined for regular grids.
    • Major bottleneck : how to  define a localized graph filters which are efficient to evaluate and learn.

Main contributions as the following:

• • Spectral formulation. If you have the ability ,you can read this paper at first.——《 The Emerging Field of Signal Processing on Graphs: Extending High-Dimensional Data Analysis to Networks and other Irregular Domains》
    • Strictly localized fifilters. 本文提出的  spectral filters 在以中心节点的为半径的最短路径内是严格局部化的。
    • Low computational complexity.  本文提出的卷积核的复杂度是和卷积核的尺寸以及边的数量成线性关系的，另外由于大多数图稀疏的，我们可以估计地认为  |ε|≪n2|ε|≪n2|\varepsilon| \ll n^{2}   并且  |ε|=kn|ε|=kn|\varepsilon|=k n ，也就是每个节点只和  kkk  个近邻有边，这样复杂度就和节点数量成线性关系。另外本方法避免了使用 Fourier basis，因此不再需要进行昂贵的  EVD (特征值分解) 以及存储  Fourier basis（使用一个大小的矩阵）。除了数据，本方法只需要存储  Laplacian matrix 这个包含  |E||E||\mathcal{E}|  non-zero values  的  sparse 矩阵。
    • Efficient pooling. 将顶点重新排列为一个二叉树结构。