KDD'16 paper, Standford

Brief Introduction

• For now, traditional & typical node embedding technique
• Propose graph feature learning approaches
  • Inspired by paper "word2vec"
  • Adopting word embedding from "skip-diagram" idea to nodes of graph
  • Propose biased walking technique

Background

Problems in Graph

• Node classification (node-level task)
  • Node degree, centrality, clustering coefficient, graphlet
• Link prediction (link-level task)
  • Distance, local neighborhood, global neighborhood
• Graph classification (graph-level task)

Node Embedding in Graph

• A mapping of nodes to a low-dimensional space of features

  • Maximizes the likelihood of preserving network neighborhoods of nodes

    

Traditionals → node embeddings → GNNs → Knowledge graphs → Deep Generative graph models

Conventional method (Linear algebraic approach)

PCA (Principal Component Analysis)

주성분 분석(PCA)

SVM (Support Vector Machine)

[Data Mining] Support Vector Machine

Shallow encoding