图机器学习基础知识——CS224W（15-deep-generation）

CS224W: Machine Learning with Graphs

Stanford / Winter 2021

15-deep-generation

Machine Learning for Graph Generation

Machine Learning for Graph Generation

• Tasks

  • Task 1: Realistic graph generation

    • Generate graphs that are similar to a given set of graphs

  • Task 2: Goal-directed graph generation

    • Generate graphs that optimize given objectives/constraints

    • E.g., Drug molecule generation/optimization

• Key Insight

  

  

  

  

GraphRNN: Generating Realistic Graphs

Paper : GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models

GraphRNN: Generating Realistic Graphs

• Key Idea

  • 使用一个节点和边的生成序列$S^{\pi }$逐步添加节点和边

    • Node-level: add nodes, one at a time

      

    • Edge-level: add edges between existing nodes

      

    

  • 将图生成问题转换成了序列生成问题

    • Node-level sequence

    • Edge-level sequence

• Architecture

  • GraphRNN has a node-level RNN and an edge-level RNN

    

    

  • Training

    

    

• Issue

  • 前述框架每生成一个节点都需要使用EdgeRNN考虑新生成的节点与前面所有节点的连接关系，生成序列过长

  • 前述框架NodeRNN生成节点是随机顺序的，现在改成以BFS node ordering，这样只需考虑当前节点与前两个节点的连接关系

    

    

    

• 如何对比生成图与原图的相似度？

  • 使用Earth Mover Distance（EMD）衡量两张图的节点度分布相似性

  • 使用Maximum Mean Discrepancy（MMD）衡量两个集合中元素的相似度（若元素为标量，则元素间相似度用L2距离之类的衡量；若元素为分布，则使用EMD衡量）

    

GCPN: Goal-Directed Graph Generation

Graph Convolutional Policy Network

• Our Goal

  • Optimize a given objective (High scores)

    • e.g., drug-likeness

  • Obey underlying rules (Valid)

    • e.g., chemical validity rules

  • Are learned from examples (Realistic)

    • Realistic Graph Generation

• key Idea

  • 这种目标导向的任务适合用强化学习建模，直接以目标作为reward

    

• Overview of GCPN

  • Insert nodes

  • Use GNN to predict which nodes to connect

  • Take action (check chemical validity)

  • Compute reward

  

• Reward