num_neighbors

[kuliev-sd]

Now parameter num_neighbors is fixed in cool_graph/config/training/default.yaml:

  - 25
  - 25

or in cool_graph/config/training/in_memory_data.yaml:

  - 25
  - 15

But these parameters are not optimal for all datasets.
It is better to have flexible algorithm for sampling neighbors.
For example, we could limit the number of neighbors by 95%-quantile of node degree, or less for 2-level neighbors