pykeen / pykeen

🤖 A Python library for learning and evaluating knowledge graph embeddings
https://pykeen.readthedocs.io/en/stable/
MIT License
1.66k stars 187 forks source link

Add reasonable defaults for NSSA Loss #25

Closed cthoyt closed 4 years ago

cthoyt commented 4 years ago

Right now, the margin and adversarial_temperature parameters do not have a default for NSSA loss. We need to pick reasonable ones and set it.

mberr commented 4 years ago

From https://arxiv.org/abs/1902.10197, where this loss was proposed, they optimize these hyperparameters in the following ranges:

The chosen parameters by the authors for FB15k, FB15k-237, WN18, WN18RR can be found at https://github.com/pykeen/pykeen/tree/master/src/pykeen/experiments/rotate :

dataset alpha gamma
FB15k 1.0 24
FB15k-237 1.0 9
WN18 0.5 12
WN18RR 0.5 6

So any of these configurations may be a good choice for a default value.