Composition-based Multi-Relational Graph Convolutional Networks

Shikhar Vashishth, et.al. “Composition-based Multi-Relational Graph Convolutional Networks”, ICLR 2020

1 简介

• 姓名：CompGCN
• 流派：relational graph embedding
• 方法：introduce relation embedding into GCN layer
  • build graph: introduce “inverse relation” and self-loop
  • node update: $h_v^{k+1} = f\bigg(\sum_{(u,r)\in\mathcal{N}(v)} W^k_{\lambda(r)} \phi (h_u^k, h_r^k)\bigg)$
  • rel update: $h_r^{k+1} = W_{\text{rel}}^k h_r^k$
  • where $W_{\lambda(r)}$ is a relation-type specific parameter, different for rel / inverse-rel / self-loop
  • trick: the dimention of initial relation representation is reduced into {relation_num}, by expressing it as a linear combination of a set of basis vectors

2 思考

Some relational GCNs(as far as I know):

• RGCN: $\sum(W_r h_i)$
• add relational weights into adjacent matrix
• CompGCN: $\sum(W_\lambda [h_i; h_r])$