dataset.py inverse triples

JuliaGast / TGB2

Temporal Graph Benchmark project repo

1 stars 0 forks source link

dataset.py inverse triples #11

Closed JuliaGast closed 5 months ago

JuliaGast commented 5 months ago

[ ] create method that creates an inverse triple for each triple in triples. inverse triple swaps subject and objsect, and increase relation id by num_rels

JuliaGast commented 5 months ago

[ ] We also need the inverse triples before we generate the negative samples -> to create negative samples for queries (d, ?, t, inv_edge_type) where we have all potential negatives for subject prediction.

for example: for tkg_negative_generator.py add the arginverse_triples_flag (default: True) if inverse_triples_flag add negatives for both directions

shenyangHuang commented 5 months ago

Given that for all tkg datasets, we need to add inverse relations, I think we should bake this directly into the dataloader. This means that the edges retrieved from dataloader will contain already the inverted relations and for the negative sampler, it would just function as normal (to generate for both directions)

shenyangHuang commented 5 months ago

This has been implemented in PR #14