Questions regarding the task definition

[moritzblum]

Hi, I was reading the paper about the dataset and was looking at the dataset in detail. Unfortunately, there is some uncertainty hindering me to fully understanding this task.

Your inductive LP task is defined as "link prediction between unseen entities - for the unseen entities, an inference graph is provided s.t. there is some information for them to work with."

The train graph is available during training, and during evaluation/test, the inference graph is available. Or is the training graph still available?

As your baseline model uses an entity's edge types, extracting features from the inference graph seems ok. Is it also allowed to extract patterns, too? Where is the border - what is allowed with the inference graph and what is not?

Moreover, according to the task definition, would it be allowed to train a new model on the inference graph? This outperforms some of the baselines...

[migalkin]

Hi, thanks for the questions, let's clarify those:

The train graph is available during training, and during evaluation/test, the inference graph is available. Or is the training graph still available?

At validation/test time, the entities between which we predict links come from the inference graph. The training graph is technically available, but it has no influence on the inference graph because they are disconnected. Whatever model was trained on the training graph, should be evaluated on the inference graph without any additional training/optimization.

Is it also allowed to extract patterns, too? Where is the border - what is allowed with the inference graph and what is not?

Yes, mining some unsupervised features is ok - you can go as far as to cellular complexes. The main restriction (used everywhere in the literature for comparability) is that you can't train/fine-tune any model on the inference graph - a model sees it for the first time at validation/test and should produce answers based on the trained weights.

would it be allowed to train a new model on the inference graph?

No, it is not allowed