optional reciprical relations

[adelmemariani]

Currently, reciprical relations are applied based on the scoring_technique. However, recipricals are not always useful. For example in Section 2 of this paper, it is stated that "The use of reciprical relations may also lead to better model performance e.g., for relations in which one direction is easier to predict)"

For example, in case of UMLS, and DistMult:

With recipricals:
{'H@1': 0.5763993948562783, 'H@3': 0.7473524962178517, 'H@10': 0.9190620272314675, 'MRR': 0.6878625225735597}

Without recipricals:
{'H@1': 0.7095310136157338, 'H@3': 0.8245083207261724, 'H@10': 0.9500756429652042, 'MRR': 0.7877015541524235}

I suggest using the "reciprical relations" as an argument for training to make it optional.

[Demirrr]

Such triples are only added if 1vsAll, kvsAll scoring techniques. Using these techniques often leads to better generalization (provided that the model has enough parameters) and reduces the testing times. This has been shown in many research papers in the last years. So, I would be against making it optional for the aforementioned scoring techniques.

[Demirrr]

That being said, we can very well implement a version of 1vsAll, KvsAll, or AllvsAll that is not coupled with reciprocal triples 😊

Let me know if you would like me to implement it😊

[adelmemariani]

I would be also happy to implement it, but I don't know how much effort it needs. Currently, I switched off recipricals for my experiments, but I will switch it on again, to see their effects... 😊

[Demirrr]

No worries. We can do pair programming in January to accelerate the process of onboarding you into the framework 😊 It should take more than an hour :)