GraphAF: A Flow-Based Autoregressive Model for Molecular Graph Generation

GraphAF: A Flow-Based Autoregressive Model for Molecular Graph Generation

Proposed Method

training: 1. extract $$A, X$$ from graph 2. $$zi = X_i + u$$ 3. model the mean and variance using GNN and MLP 4. $$\epsilon_i = (z_i - \mu_i) \odot \frac{1}{\alpha_i}$$ 5. calculate the likelihood $$L_i = -\log (Prod(p\epsilon(\epsilon_i))) - \log(Prod(\frac{1}{\alpha_i}))$$ 6. Once the node is fixed, do the similar things for edges

test: 1. sample $$\epsiloni$$ and $$\epsilon{ij}$$ from normal distribution 2. $$z_i = \epsilon_i \odot \alpha_i + \mu_i$$, similarly for edges