🪑 Benchmark PyKEEN's negative samplers' false negative rates. Run with:
$ python -m pip install -r requirements.txt
$ python main.py
It will keep track of your username via getpass
and the git hash if you
have a development version of PyKEEN installed.
The plot demonstrates that the pseudo-typed negative sampler should be used in combination with filtering. The Nations dataset has a unique pattern because it was constructed under the closed world assumption (unlike most knowledge graphs) and therefore most negative samples are false negatives.