CompGCN

COMPOSITION-BASED MULTI-RELATIONAL GRAPH CONVOLUTIONAL NETWORKS

1 INTRODUCTION

原来的CNN, RNN等方法不能直接应用到graph上,因此最近GCN被提出来了。

但是初始的GCN方法主要集中与无向图,最近的针对有向图的方法类如R-GCN,存在over-parameterization问题。

there is a need for a framework which can utilize KG embedding techniques for learning task-specific node and relation embeddings.

COMPGCN addresses the shortcomings of previously proposed GCN models by jointly learning vector representations for both nodes and relations in the graph

两个方面叙述

  • GCN: 原始的GCN,之后的各种拓展,都在MPNN框架下,本论文提出的也是这样,但是专门为relational data设计过
  • KE: translational、semantic matching based、neural network based

3 BACKGROUND

GCN on undirected graph

图的表示形式: \[ G=(\cal{V},\cal{E},\cal{X}) \] 其中\(\cal{X}\)表示所有entity的初始feature,\(\cal{X}\in \cal{R}^{ |\cal{V}|\times d_0 }\)

获取归一化后的self-connection邻接矩阵: \[ \hat{A}=\tilde{D}^{-\frac{1}{2}}(A+I)\tilde{D}^{-\frac{1}{2}} \\ \tilde{D}_{ii}=\sum_j{(A+I)_{ij}} \] 某一层的GCN: \[ H^{k+1}=f(\hat{A}H^kW^k) \\ H^0=\cal{X} \]

GCN on directed graph

图的表示形式: \[ G=(\cal{V},\cal{R},\cal{E},\cal{X}) \] \(\cal{R}\)表示relation的集合。

这种情况下对于关系数据的处理就存在区别了,基于Encoding sentences with graph convolutional networks for semantic role labeling 中提出的假设,

information in a directed edge flows along both directions

因此构造出反向关系inverse relation: \[ (u,v,r)\in \cal{E}\ \ and\ \ (v,u,r^{-1})\in \cal{E^{-1}} \] 此时的GCN: \[ H^{k+1}=f(\hat{A} H^k W^k_r) \\ H^0=\cal{X} \]

4 CompGCN DETAILS

图: \[ G=(\cal{V},\cal{R},\cal{E},\cal{X},\cal{Z}) \\ \cal{X}\in \cal{R}^{ |\cal{V}|\times d_0 } \\ \cal{Z}\in \cal{R}^{ |\cal{R}|\times d_0 } \] 构造关系: \[ \cal{E^{'}} = \cal{E}\ \cup\ \{ (v,u,r^{-1}) | (u,v,r)\in \cal{E}\ \}\ \cup\ \{ (u,u,T) | u\in \cal{V} \} \] embedding更新方式: \[ h_v^{k+1}=f(\sum_{(u,r)\in \cal{N}_v} W_{\lambda(x)^k}\phi(x_u^k, z_r^k)) \] 其中\(\phi\)函数,为了减少参数,可以为下面的三种方式,当然可以拓展为更多的方式: \[ Sub: \ \phi(x_u, z_r) = x_u - z_r \\ Mult:\ \phi(x_u, z_r) = x_u * z_r \\ Circular-correlation:\ \phi(x_u, z_r) = x_u \star z_r \] 其中的关系权值矩阵: \[ W_{dir(r)}= \begin{cases} W_o,\ r\in \cal{R} \\ W_i,\ r\in \cal{R}_{inv} \\ W_S\ r\in \cal{T}(self-loop) \end{cases} \] 对于关系relation的处理,与KBGAT一样: \[ z_r^{k+1} = W_{rel}z_r^k \\ W_{rel}\in R^{d_1\times d_0} \] 在第一层初始的时候,对于relation的定义是bias-vector。 \[ Z_r = \sum_b^B \alpha_{br}\bold{v}_b \\ \{ \bold{v}_1, \bold{v}_2,\cdots \bold{v}_B \} \]

5 EXPERIMENTAL SETUP

进行了下面三个任务:

  • Link Prediction:FB15k-237,WN18RR
  • Node Classification:MUTAG (Node) , AM
  • Graph Classification:bioinformatics dataset:MUTAG (Graph) , PTC

6 RESULTS

研究了下面四个方面的问题:

  1. 在link prediction上的效果
  2. 选择不同的composite operation效果
  3. 模型对于不同数量的relation的数据集的效果
  4. 在node和graph classification的效果

We find that with DistMult score function, multiplication operator (Mult) gives the best performance while with ConvE, circular-correlation surpasses all other operators.

具体结果略