CS224W Lecture 3 Node Embeddings

Introduction

  • traditional ml for graphs
    • input-graph => feature-engineering => structured feature => learning algorithm => prediction
  • graph representation learning
    • input-graph => feature-engineering representation learning => structured feature => learning algorithm => prediction
    • goal: efficient task-independent feature learning
    • task: map nodes into an embedding space
      • possible downstream tasks: node classification, link prediction, graph classification, anomalous node detection, clustering, …

node embeddings: encoder and decoder

  • encoder: maps from nodes -> embeddings
  • decoder(similarity function): maps embeddings -> similarity scores
  • learning node embeddings: optimize parameters of the encoder
  • simplest approach: encoder is just an embedding-lookup (DeepWalk, node2vec)
  • deep encoders -> lecture 6
  • objective: maximize similarity score for node pairs that are similar
  • key choice: how can we define node similarity

random walk approaches for node embeddings

random walk

  • similarity score approximates a probability that two nodes co-occur on a random walk over the graph
  • why random walk? why?

    • random walk optimization is computationally expensive => use negative sampling
    • how to solve optimization problem? => SGD
  • strategy to walk randomly
    • simplest: just fixed-length, unbiased random walk -> DeepWalk
    • issue: similarity is too constrained
    • how can we generalize this? -> node2vec

Node2Vec

  • goal: embed nodes with similar network neighborhoods close in the feature space
  • key observation: flexible notion of network neighborhood leads to rich node embeddings node2vec

    • BFS strategy can capture local features
    • DFS strategy can capture global features
  • hyperparameter for node2vec
    • p: return back to the previous node
    • q: in-out parameter, ratio of BFS vs DFS
  • biased 2nd-order random walks
    • idea: remember where the walk came from biased random walk
  • node2vec algorithm

    1. compute random walk prob
    2. simulate random walks of specific length starting from each node
    3. optimize the node2vec objective using SGD

embedding entire graphs

  • goal: embed subgraph or entire graph
  • approaches
    • (simplest) average the node embeddings
    • add virtual node to represent (sub)graph and run standard graph embedding technique virtual node
    • anonymous walk embeddings awe

      • sample use of anonymous walk aw
    • learn walk embeddings of anonymous walk lwe1 lwe2
    • hierarchiccal embeddings -> later in this lecture
      • we can hierarchically cluster nodes in graphs, and sum/avg the node embeddings
November 11, 2021
Tags: cs224w