While ranking training seems like to work now (except timing info :( ), for 1000 retrieved events I have 5 likes (rbr.io is retrieving lots more events than other clients because it shows all parent comments for now).

I guess the solution will be to get liked events as well from the user's relays to get liked / non-liked training data ratio closer to 50% then reweight the liked events to have very small, but much more learnable weights. For example if 5 like events are exchanged to 100, the liked data should have 0.05 item weighting when training.

Of course a modified Hessian for log likelihood needs to be computed for Newton-Ralph method for finding max log likelihood, so if somebody is good at math, they are welcome to help adding item weights to the computation :)