论文笔记:Spectral Networks and Deep Locally Connected Networks on Graphs

这篇论文发表在2014 ICLR会议,是第一篇研究将CNN泛化到非欧式空间的论文,参考资料1直接称之为“第一代GCN”。

 

主要贡献:

将CNN泛化到非欧几里得空间,并提出两种并列的model,分别是基于spatial和基于spectral。

 

1. Spatial Construction:

进行K层聚类,每一层中又会有若干个“filter”。每层与每层之间的神经元数目是通过聚类而成,将上一层的聚类结果对应为下一层的神经元。如图,第0层有12个节点,经过聚类,第1层有6个节点,第2层仅有3个节点。

Figure 1 展示了节点聚类的过程。

论文笔记:Spectral Networks and Deep Locally Connected Networks on Graphs_第1张图片

Figure 2 展示了卷积 和 池化的整个过程。

论文笔记:Spectral Networks and Deep Locally Connected Networks on Graphs_第2张图片

论文笔记:Spectral Networks and Deep Locally Connected Networks on Graphs_第3张图片

公式2.1 是本文提出的第一类model的卷积公式,Lk是池化,h是激活函数,F为每一层的每一个特征的权值向量。

(此处可以联系到欧式空间中CNN的卷积,池化)

 

2. Spectral Construction

基于图谱理论来阐述,有关graph spectral的相关知识,见参考资料1。

卷积公式如下所示:

论文笔记:Spectral Networks and Deep Locally Connected Networks on Graphs_第4张图片

 

 

3.扩展补充

spectral methods 和 spatial methods区别:

1.Spectral methods usually handle the whole graph simultaneously and are difficult to parallel or scale to large graphs.

2.Spatial methods directly perform the convolution in the graph domain by aggregating the neighbor nodes’ information. Together with sampling strategies, the computation can be performed in a batch of nodes instead of the whole graph, which has the potential to improve the efficiency

 

参考资料:

1.从CNN到GCN的联系与区别——GCN从入门到精(fang)通(qi)
https://www.zhihu.com/question/54504471

2.论文笔记之Spectral Networks and Deep Locally Connected Networks on Graphs - BVL的博客 - CSDN博客
https://blog.csdn.net/bvl10101111/article/details/53426226

你可能感兴趣的:(GCN)