Information
title：
Masked Graph Convolutional Network
authors：
Liang Yang, Fan Wu, Yingkui Wang, Junhua Gu and Yuanfang Guo
Abstract
Semi-supervised classification is a fundamental technology to process the structured and unstruc- tured data in machine learning field. The tradi- tional attribute-graph based semi-supervised clas- sification methods propagate labels over the graph which is usually constructed from the data fea- tures, while the graph convolutional neural net- works smooth the node attributes, i.e., propagate the attributes, over the real graph topology. In this paper, they are interpreted from the perspective of propagation, and accordingly categorized into sym- metric and asymmetric propagation based methods. From the perspective of propagation, both the tradi- tional and network based methods are propagating certain objects over the graph. However, different from the label propagation, the intuition “the con- nected data samples tend to be similar in terms of the attributes”, in attribute propagation is only par- tially valid. Therefore, a masked graph convolu- tion network (Masked GCN) is proposed by only propagating a certain portion of the attributes to the neighbours according to a masking indicator, which is learned for each node by jointly consid- ering the attribute distributions in local neighbour- hoods and the impact on the classification results. Extensive experiments on transductive and induc- tive node classification tasks have demonstrated the superiority of the proposed method.
Core Issue
不同于 label propagation，就属性而言，相连的节点倾向于是相似的，这一直觉在属性传播中只是部分合法的。作者提出 masked graph convolution network，根据一个属性指示器（attributes indicator）来决定一个节点要将它的哪些属性传给邻居。考虑到在局部邻居的属性分布和对分类结果的影响，为每一个节点学习一个属性指示器。
传统GCNNs和Masked GCN 的比较如下：
其中黑色代表值较大。
Methods
目标函数如下:
$h_i$是节点$v_i$的属性,$w_{i,j}$由公式15定义。
根据目标函数，属性更新公式如下：
mask 的定义如下：
全连接预测层如下:
Innovation
考虑到在局部邻居的属性分布和对分类结果的影响，为每一个节点学习一个属性指示器。据一个属性指示器（attributes indicator）来决定一个节点要将它的哪些属性传给邻居。
Future Work
In the future, our Masked GCN will be extended to other heterogeneous networks
ps:可考虑为不同类型的邻居学习一个mask indicator。
Shortcomings
为每一个节点学得的属性指示器，是一对多关系，可能存在这样的关系，它的某个属性需要传给部分邻居，不传给另一部分邻居。虽然有考虑邻居权重，但依然会存在平均问题（少数服从多数）。
PS:但如果学习一个一对一的属性指示器，则可能会令信息太细粒，导致更高阶的结构信息不能很好的抓住。