I think, for thgl-github, the negative sampling is not great
Reason:
the recurrency baseline should not be good for this dataset. It has a rather low recurrency degree, and applying the relaxed recurrency baseline only seems to lead to not good results either
Still, it gets a rather ok mrr of 0.17
In the examples that I checked, for a node-relation combi, there have been 2000-6000 different objects
-> 1000 is a small subsample of the possible nodes
Given that this is our smallest dataset, I strongly suggest that we do not do any negative sampling here.
We have a very small recurrency degree. I expect this dataset will be very hard. hm.
I think, for thgl-github, the negative sampling is not great Reason: the recurrency baseline should not be good for this dataset. It has a rather low recurrency degree, and applying the relaxed recurrency baseline only seems to lead to not good results either Still, it gets a rather ok mrr of 0.17 In the examples that I checked, for a node-relation combi, there have been 2000-6000 different objects -> 1000 is a small subsample of the possible nodes Given that this is our smallest dataset, I strongly suggest that we do not do any negative sampling here. We have a very small recurrency degree. I expect this dataset will be very hard. hm.