From Graph to Knowledge Graph – Algorithms and Applications [Note2]

Part3. Graph Representation Learning

Problem: structural feature engineering over the graph is expensive.
=> Graph Representation Learning aimed to learn latent(潜在的) feature matrix

Representation learning for graph mining?
The goal is to map each node into a latent low-dimensional space such that network structure information is encoded into distributional node representations.

1. Skip-gram based graph embedding

skip-gram based word embedding

Consider sentence as a node path, each node representing a word.


Two steps:

  1. random walks over graph to generate node path.
  2. skip gram model to apply to node path generated in the first step.

Skip gram with negative sampling

  • For sufficiently large dimension d, the objective of SGNS(Skip-gram with negative sampling) is equivalent to factorizing the PMI matrix(Pointwise mutual information matrix).

2. Heterogeneous graph embedding

3. Graph convolutional network

Ref: From Graph to Knowledge Graph – Algorithms and Applications