GNN algorithm(5): Graph Encoder-Decoder Mechanism

介绍Graph Representative Learning的 Encoder-Decoder框架。

2017: Representation learning on graphs: Methods and applications

Problem Definition

• Graph机器学习的核心问题：如何把Graph结构信息纳入机器学习模型中。常用的Graph结构信息包括：节点的全局位置(Global Position)和节点的局部近邻结构(The structure of the node's local graph neighborhood)
• Graph Embedding目标：学习node或entire (sub)graph的低维embedding，使得embedding space中的几何关系能够反应原始Graph的结构信息，如两个node节点在embedding space中的距离能够反应原始高维空间Graph中二者的相似性。(Learn embeddings that encode graph structure)