find best KG embedding model - pyKEEN

[kdutia]

Use a sample of data; test TransE and RotateE models.

Create data

• [x] choose subset of properties for graph embeddings (i.e. not literals, unless they are categorical e.g. occupations)
• [x] create processed set of triples with this subset of properties
• [x] create sensible-sized samples of all the triples for training and test sets (for cost-effective experimentation!)

Run hyperparameter optimisation

• [x] choose hyperparameters to vary and their ranges
• [ ] set up and run hyperparameter optimisation
• [ ] train models on best parameter configurations for TransE and RotatE

Evaluate For the best TransE configuration and the best RotatE configuration:

• [ ] what does a plot look like?
• [ ] what do the K nearest neighbours look like for a sample of entities?
• [ ] how does a linear classifier trained to predict record type perform on the embeddings? (like here)

[kdutia]

not using pykeen - see #309 and #310