Exploration of Graph Network Variants

[Simsso]

1. Train a screenshot feature extractor and store its weights. Its output is e.g. a 32 dimensional vector.
2. Train several graph networks that use its weight and experiment with different configurations. The low number of parameters should allow for fast iteration and many experiments.

Things to try:

• [ ] Weight sharing between consecutive GN blocks
• [ ] Fully-connected vs. original graphs
• [ ] Sum instead of avg aggregation
• [ ] Conversion of links into two edges, pointing both ways

[Simsso]

Experiments

• [x] GN avg ([baseline+avg])
• [x] GN max ([baseline+max])
• [x] GN 1 core
• [ ] GN 3 core
• [ ] GN 3 core (with weight sharing)

Weighting

• [ ] w_b = 1
• [ ] b = 1
• [ ] b = 2
• [ ] b = 10
• [ ] b = 100

Edges

• [ ] no edges
• [ ] default
• [ ] bi-directional
• [ ] full

[Simsso]

Variants to try are set. Keeping track of the experiments in this new wiki page.