Variational Graph Autoencoder

[max-acc]

Research if the usage of a VGAE is viable.

Pipeline

Hit Features + Graph -> Encoder -> Latent z_i -> Decoder -> probabiltiy(edge | z_i, z_j)

This approach would reconstruct the adjacency matrix directly.

Possible Limitations

• High class imbalance
• Needs Graph Pruning before Message Passing
• Posterior Collapse: probably same issue as with images

Research Directions

• Geometric invariances may be forceable, so that rotation and translation would not change the structure