metapath2vec: Scalable Representation Learning for Heterogeneous Networks

Time:

 

Author: Yuxiao Dong, Nitesh V. Chawla, Anathram Swami

 

• homogeneous networks: representation of singular type of nodes and relationships, such as DeepWalk, LINE, and node2vec
• heterogeneous networks: representation of diverse node types and/or relationships between nodes

 

Abstract

• metapath2vec model
  • 基于meta-path的随机游走形成node的heterogeneous neighborhood
  • 再使用heterogeneous skip-gram model形成node embeddings
• metapath2vec++
  • 使用heterogeneous negative sampling-based方法，在heterogeneous networks里使用了structural和sematic correlations，来预测一个节点的heterogeneous neighborhood

Introduction

• 和传统的meta-path-based methods相比，使用低维隐向量，可以很好的计算出没有meta-path连接的节点之间的相似度
• 目标：学习不同类型的节点的低维隐向量表示，使得保留heterogeneous network的结构和语义的概率最大化
• 优势：
  • 在node cla