Add novel link prediction pipeline about pykeen HOT 2 CLOSED

Question: do models trained under LCWA or sLCWA have to behave differently here?

The behavior should be the same for both approaches. But the user should be aware that if the model was trained based on the LCWA training approach (in combination with 1:N-scoring), that for each (h,r)-pair for which a triple exists in the KG, all (h,r,t_i) that are not part of the KG have been considered as false examples during training. Therefore, asking the model to provide predictions for such (h,r)-pairs might not be very useful.

In general, we should just make sure to call the functions model.predict_scores_all_heads()/model.predict_scores_all_tails(), and not directly model.score_h()/model.score_t(), because the former make sure that the model uses the appropriate scoring functions in case it was trained with inverse relations.

from pykeen.

It should be clear if higher score or lower score means more likely to be a real edge.

In our framework, we always treat higher scores as higher likelihood.

We might want to have it automatically filter out entities corresponding to edges already in the original knowledge graph.

👍

Question: do models trained under LCWA or sLCWA have to behave differently here? Does this functionality have to behave differently based on the loss function?

For point-wise loss functions, the score are calibrated across different (h, r) / (r, t)-pairs (since we have a fixed "target value" for each triple during training). For pairwise loss functions, this is not the case, and during training we only imposed a meaningful order between tail entities for the same (h,r) (and vice-versa).

from pykeen.