Removing stochasticity during evaluation

[anton-sturluson]

In a stochastic edge prediction training setting, I realized a different set of neighbors is sampled for each edge due to stochasticity, i.e., the g.sample_neighbors method returns different neighbors each time it is run. And this is causing quite a bit of variance in the evaluation result since some regions in my graph have high connectivity.

1. Is there a way to fix the neighbors sampled by the node sampler? I tried to play around with seed in vein.
2. If 1) is impossible, what’s a good way to fix this issue? I can think of oversampling, since different topologies will be selected for the same edge, and the model will be exposed to multiple views.
3. Any reference to papers dealing with such issues?

[BarclayII]

You can try setting dgl.random.seed().