My datasets problems about dgfraud

[YingtongDou]

1. You can keep only a few informative neighbors under each relation. I don't know if such sampling could efficiently reduce the node size. Anyway, your sampling method should reduce the graph size while keeping enough neighbors for each node under each relation.

2. If you do not know how different encoding methods contribute to your classification task, you can test them using feature selection tools from sklearn and keep the simple and useful one. Or you can test them directly with GEM.