GMT【公式整理】：Accurate Learning of Graph Representations with Graph Multiset Pooling

Overall Architecture

获得embedding H

Pooling

1. Graph Multiset Pooling (GMPool)
2. Self-Attention Block (SelfAtt)
3. GMPool with k = 1
   由接下来几个部分逐一展开解释

Graph Multi-head Attention (MH)

standard multi-head attention

use GNN to generate Key and Value：preserve graph structure

Graph Multiset Pooling with Graph Multi-head Attention (GMPool)

parameterized seed matrix S as Query
interactions between k seed vectors (queries) in S and n nodes (keys) in H

Self-Attention for Inter-node Relationship (SelfAtt)

consider relationships between nodes (Ablation证明这部分非常重要)