HAN

Heterogeneous Graph Attention Network

2019-4-13

1 INTRODUCTION

HAN设计了两个层次的attention机制,

  • Semantic-level attention:在不同的meta-path中选择weight
  • Node-level attention:对于一个节点,它的邻居的weight

The graph convolutional neural work generally falls into two categories, namely spectral domain and non-spectral domain.

4 THE PROPOSED MODEL

总的来说,HAN包括了两个层次的attention,node-level和semantic-level,node-level attention的输出作为semantic-level层次的输入。

4.1 Node-level Attention

对于\((i,j,\Phi)\)\(i,j\)表示节点,\(\Phi\)表示meta-path。

node-level attention针对的目标是同一个meta-path下的nodes的weight。

首先根据node type确定投影embedding, \[ h^{'}_i=M_{\phi_i}h_i \\ \phi_i: node \ i \ type \] 计算attention value, \[ \alpha_{ij}=\frac{exp(\sigma(a^T[h_i^{'}||h_j^{'}]))}{\sum_{k\in N_i^{\Phi}} exp(\sigma(a^T[h_i^{'}||h_k^{'}]))} \] 把所有node的embedding结合起来, \[ z_i^\Phi = \sigma(\sum_{j\in N_i^{\Phi}} \alpha_{ij}^\Phi h_{ij}^{'} ) \] 类似于GAT,采用multi-head attention, \[ z_i^\Phi = ||_{k=1}^K \sigma(\sum_{j\in N_i^{\Phi}} \alpha_{ij}^\Phi h_{ij}^{'} ) \] 这是一种meta-path下一个节点\(i\)的最终输出,对于所有的\(\Phi\)与全部的node,产生\(\{z_i^{\Phi_0}, z_i^{\Phi_1},\dots z_i^{\Phi_p}\}\)

4.2 Semantic-level Attention

要计算各种类型的meta-path的weight,就要在全局的情况下计算, \[ \omega_{\Phi_p}=\frac{1}{|V|}\sum_{i\in V}q^T tanh(Wz_i^{\Phi_p}+b) \\ \beta_{\Phi_p}=softmax(\omega_{\Phi_p}) \] 最后求和,得到最终的embedding, \[ Z_i=\sum_{p=1}^P \beta_{\Phi_p}z_i^{\Phi_p} \] image-20200221215000895

5 EXPERIMENTS