also, federated training is a possibility but makes everything a bit more complicated. i guess the most important question is: what features turn out to be predictive. maybe the protocol itself can be leveraged more (network/metadata).