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.
Consider sentence as a node path, each node representing a word.
- random walks over graph to generate node path.
- 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).