Inductive Datasets

[vthost]

Hi, in the paper it says the following, but we observed several overlapping entities (e.g., /m/080knyg in fb237_v1_ind). Are we maybe missing steps in the data processing, or could you please detail what you mean by inductive setting?

"F. Inductive Graph Generation
The inductive train and test graphs examined in this paper do not have overlapping entities."

Thank you already!

[kkteru]

Hi!

Quoting from the paper:
Each of our datasets consists of a pair of graphs: train-graph (fb237_v1) and ind-test-graph (fb237_v1_ind). These two graphs (i) have a disjoint set of entities, and (ii) train-graph contains all the relations present in ind-test-graph.

To elaborate, there will be no overlapping entities between fb237_v1 and fb237_v1_ind. The models are trained on fb237_v1 and tested on fb237_v1_ind in the inductive setting. Following is an illustrative example:

So, the model wouldn't have seen S. Curry, Ayesha, and CA during training. But an inductive model should still be able to make an accurate prediction by learning the relational rule from the train-graph. Hope that clarifies your question.

This is in fact a rather confusing point. We will try to clarify further in the paper. Thanks for your question!

[vthost]

Thank you for the fast reply!
It now makes sense. Maybe you should not call edges from the inductive test graph train edges. To me, it seems that they are not used to train the model, but rather additional input at test time.