Open nomisto opened 5 months ago
In this particular dataset, InductiveFB15k237
, the validation and test graphs are guaranteed to be disjoint - so we don't need to filter for them; however, it would not hurt much and might be a good adjustment to make sure you do not forget to do this filtering in other datasets that do not have this guarantee.
I don't think that is true. The training and inference graph are disjoint. But not the test and validation set of the inference graph. E.g. test.txt of fb237_v1_ind contains the triple /m/0p9xd /music/genre/artists /m/01304j While valid.txt of fb237_v1_ind contains the triple /m/02x8m /music/genre/artists /m/01304j Now for a query genre_artists(?, /m/01304j) the validation triple /m/02x8m would not get filtered
Is there a reason you omit the validation triples of the inference (test-ind) graph from filtering in your example here when doing evaluation on the test set?
I would expect that the test evaluator would also include the validation triples for filtering, like