Graph Auto-Encoders (GAEs)

https://www.youtube.com/watch?v=qA6U4nIK62E

• like autoencoders, their goal is to produce a matrix representation of the input
  • $\bar{X}=GCN(A,X)=ReLU(\bar{A}X{W}_0)$
    • $\bar{X}$ is the representation
  • Note that $\bar{A}$ is the normalized adj matrix:
    • $\bar{A}=D^{-1/2}A{D}^{-1/2}$
• 
• how to reconstruct the input graph?
  • Solution 1:
    • Perform an inner product all indexes in the latent space (and sigmoid the result) to get the adj matrix back
      • 
    • why does this work?

• variational autoencoders