MatKG

[vinven7]

Describe the benchmark KG dataset:

This is a dataset of materials, their applications, properties, characterization methods, synthesis methods, descriptors, and symmetry labels extracted from over 5 Million scientific publications in material science. It has over 6 million triples covering 70,000 entities.

E.g., what domain does this benchmark KG cover? Are there any special parts besides triples?

The domain is Material Science. Each triple has another field called 'count', which is a numerical parameter that is a weight. It is the number of documents in which the triples appear.

Dataset is pre-stratified into train/test/valid [ ]

NO

Publication(s) or website describing benchmark KG dataset:
https://www.nature.com/articles/s41597-024-03039-z.pdf

URL for downloading the benchmark KG dataset:
https://zenodo.org/records/10144972

[mberr]

Hi @vinven7 ,

There is currently no support for triple weights, although there has been interest in adding support for this, see #1142.

As a temporary workaround, it may be possible to add a user-defined count threshold to the KG dataset that would discard triples below a certain count. This way, the resulting KG would be in a format compatible with all existing structures.

We would be happy to accept a PR, especially since you seem to be involved in the creation of this KG. If you would like to participate in adding support for triple weights, we would also appreciate your contribution, but this seems to be a larger effort than the baseline I suggested above.