Novel Link Prediction

After training, the interaction model (e.g., TransE, ConvE, RotatE) can assign a score to an arbitrary triple, whether it appeared during training, testing, or not. In PyKEEN, each is implemented such that the higher the score (or less negative the score), the more likely a triple is to be true.

However, for most models, these scores do not have obvious statistical interpretations. This has two main consequences:

1. The score for a triple from one model can not be compared to the score for that triple from another model

2. There is no a priori minimum score for a triple to be labeled as true, so predictions must be given as a prioritization by sorting a set of triples by their respective scores.

After training a model, there are four high-level interfaces for making predictions:

1. pykeen.models.predict.get_prediction_df() for a head / relation / tail scoring, given the other two

2. pykeen.models.predict.get_all_prediction_df() for prioritizing links, i.e., calculating scores for all triples

3. pykeen.models.predict.predict_triples_df() for computing scores for explicitly provided triples

Scientifically, pykeen.models.predict.get_all_prediction_df() is the most interesting in a scenario where predictions could be tested and validated experimentally.

After Training a Model

This example shows using the pykeen.pipeline.pipeline() to train a model which will already be in memory. Each of the high-level interfaces are exposed through the model:

>>> from pykeen.pipeline import pipeline
>>> from pykeen.models import predict
>>> # Run the pipeline
>>> result = pipeline(dataset="Nations", model="RotatE")
>>> # save the model
>>> result.save_to_directory("doctests/nations_rotate")
>>> model = result.model
>>> # Predict tails
>>> predicted_tails_df = predict.get_prediction_df(
...     model=model, head_label="brazil", relation_label="intergovorgs", triples_factory=result.training,
... )
>>> # Predict relations
>>> predicted_relations_df = predict.get_prediction_df(
...     model=model, head_label="brazil", tail_label="uk", triples_factory=result.training,
... )
>>> # Predict heads
>>> predicted_heads_df = predict.get_prediction_df(
...     model=model, relation_label="conferences", tail_label="brazil", triples_factory=result.training
... )
>>> # Score all triples (memory intensive)
>>> predictions_df = predict.get_all_prediction_df(model, triples_factory=result.training)
>>> # Score top K triples (computationally expensive)
>>> top_k_predictions_df = predict.get_all_prediction_df(model, k=150, triples_factory=result.training)
>>> # Score a given list of triples
>>> score_df = predict.predict_triples_df(
...     model=model,
...     triples=[("brazil", "conferences", "uk"), ("brazil", "intergovorgs", "uk")],
...     triples_factory=result.training,
... )

Loading a Model

This example shows how to reload a previously trained model. The pykeen.pipeline.PipelineResult.save_to_directory() function makes a file named trained_model.pkl, so we will use the one from the previous example.

>>> import torch
>>> from pykeen.datasets import get_dataset
>>> model = torch.load("doctests/nations_rotate/trained_model.pkl")
>>> training = get_dataset(dataset="nations").training
>>> # Predict tails
>>> predicted_tails_df = model.get_prediction_df(
...     head_label="brazil", relation_label="intergovorgs", triples_factory=training
... )
>>> # everything else is the same as above

There’s an example model available at https://github.com/pykeen/pykeen/blob/master/notebooks/hello_world/nations_transe/trained_model.pkl from the “Hello World” notebook for you to try.

Potential Caveats

The model is trained on its ability to predict the appropriate tail for a given head/relation pair as well as its ability to predict the appropriate head for a given relation/tail pair. This means that while the model can technically predict relations between a given head/tail pair, it must be done with the caveat that it was not trained for this task.