the performance of the link prediction pipeline on the karate dataset is terrible (around 45-50%, which is the baseline), but on the usmle dataset we are doing much better than chance, which seems to be a harder problem at the first glance.
a conjecture is that there is a bug which affects the karate performance, but not the usmle. it might be the case that this is because karate has only one link type, because we are handling that differently. it also may be for other reasons (maybe karate has attributes?).
@bvarjavand seek and destroy (as in: find and fix)
the performance of the link prediction pipeline on the karate dataset is terrible (around 45-50%, which is the baseline), but on the usmle dataset we are doing much better than chance, which seems to be a harder problem at the first glance.
a conjecture is that there is a bug which affects the karate performance, but not the usmle. it might be the case that this is because karate has only one link type, because we are handling that differently. it also may be for other reasons (maybe karate has attributes?).
@bvarjavand seek and destroy (as in: find and fix)