Is it possible to include an ontology (created in for example Protoge) in a format that is useful as input for the training?
It feels like a great way to boost the accuracy, especially for smaller graphs. Take edge detection for example, if it is known that (V1 ->e1->V2) is mutually exclusive to (V1->e2->V3) then some guesses can be ignored.
If this is not currently possible - Is that for a good reason or that no one has come around to it yet? 
Are you asking whether it’s possible to do link prediction with some sort of logical constraints involved? I feel this is more a general question of how to combine deep learning methods with rule-based methods. I’m not familiar with Protege though so I can only offer some general advices.
For link prediction in particular, a deep learning model for link prediction merely gives a single score (which is just a real value) for a candidate pair of nodes to indicate the likelihood of edge existence. You can apply the logical constraints either before generating the candidate node pairs for the model, or after the model gives scores to a set of candidates to filter out the unreasonables.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.