How does it work?: This model is based on this paper.
To summarize, it uses a variational autoencoder to embed data samples into a latent space parametrized by an adjacency matrix, while minimizing the ELBO loss and maintaining a continuous acyclicity constraint.
Read more in this Medium Article my teammate and I worked on