GNN PageRank

GNN学习笔记

同济子豪兄：改变世界的谷歌PageRank算法课程笔记

Google PageRank

• The Anatomy of a Large-Scale Hypertextual Web Search Engine

Introduction

• The web as a directive graph
• Ranking nodes on the graph
• Core idea:
  • Links as vote (In-comming links)
  • Links from important pages count more
  • Recurisive question
• Solusion

  • Measure the importance of nodes in the graph using the link structure.
    

  • r_j = \Sigma_{i \leftarrow j}{\frac{r_i}{d_i}}

    • d_i is the outdegree of node i

  • Stochastic adjacency matrix M, Rank vector r

    • M_{j,i} = 1 / d_i

    • \Sigma_i{r_i} = 1

    • r \get M \cdot r

    • EigorVector(特征向量)
      

    • 收敛：r = M(M(…M(r)))

      • r_i^{t+1} - r_1^t < \epsilon

  • Random Walk

    • 模拟的思路
    • 可以理解为在directive graph中random walk，pagerank就是它可能走到这个node的概率。
    • Spider Traps
      • With prob. \beta, follow a link at random
      • With prob. (1-\beta), jump to a random page
    • Dead-ends
      • Solution: Make matrix column stochastic by always teleporting when there is nowhere else to go

  • Integrate Procedure