General data pipeline

[wagpa]

DRAFT

in: Embedding (V, E, R) where

• V are the node features
• E are the (known) edges between the nodes
• R are the embedded nodes (with lower dimension)

1. Normalize R onto a unit disc
2. For every epoch (or batch?) generate a sample R' of R (that is representative)
3. Generate node pairs of R': p = (n0_x, n0_y, n1_x, n1_y) with label being either 0 or 1 if the pair had an edge between them
4. Use a weighted sampler so that the number of 0- and 1-labeled node pairs are (expected to be) of same size
5. Train the model on the node pairs
6. Evaluate embedding by reconstructing the graph with node pairs and checking the accuracy