Should we filter out objects exist in whole dataset or just in training set?

[boldchan]

When we perform "fil" evaluation on object prediction, prediction score for some objects will not be used in ranking. For a looser evaluation, objects happen with the same (subject, relation) pair will be filtered out regardless of their timestamp. The question is what is the scope of events, in which we look for the objects of the same pair. Is the scope only training set or the whole data set?

KGReasoning/tKGR/utils.py

Line 193 in eddf7ce

def get_sp2o(self):