Do you mind sharing the code about how to get the dataset division?

[zhiweihu1103]

Nice work! Thank you very much for the open source code. In the original paper, the FB15k-237, NELL995 and WN18RR datasets are divided into V1 to V4 versions. Because of the needs of my own paper, I want to divide the FB15k dataset similarly to generate four versions of datasets such as FB15k_v1, FB15k_v1_ind, etc. if it is convenient for you, would you like to open source the code divided by different versions of the dataset? Looking forward to your reply.

[WaHaHa11122]

Hello, I am a beginner and just started to study this field. After reading your paper, I am very interested in your research. Is it convenient to share the code about data set partitioning with me?Looking forward to your reply.

[nomisto]

@kkteru Nice work! Would you like to open source the code for the creation of the datasets?

[jnaeg]

I am pretty sure it is here:

grail/utils/prepare_meta_data.py

Line 155 in 2a3dffa

if __name__ == '__main__':

If you want to have transductive sets for training, val, and testing, probably run this afterwards:

grail/utils/clean_data.py

Line 84 in 2a3dffa

if __name__ == '__main__':

Why the second step: Prepare_meta_data only checks for seen relations, not for entities. To filter entities not seen from the val/test set, run clean_data.py on your transductive split folder. See also this default parameter.

some things of note:

• all data splits are only based on the original kgs' training split.

• Not sure what 'meta data' and 'meta data dir' are doing. My hypothesis is that this legacy code/later got renamed to '_ind' instead of '_meta', as the resulting saved files (and their contents) and folder structure match what we expect: see paper "We remove the samples training graph from the whole graph and sample the test graph using the same procedure."

• as the paper only says "The parameters of the above process are adjusted to obtain a series of graphs of increasing size", I have no idea how to set them in prepare_meta_data to reproduce the different splits.

  • My first idea would be via the max-nodes-per-hop parameter, as this would also fit the very round numbers of entities reported in the paper for Fb15k.
  • Another possible way is to conjecture that the number behind the version corresponds to the number of seed nodes, as a) the default is set to 1, and b) NELL-995-v1 only consists of a star-shaped kg around one node called 'public braodcaster'... however, for more than one seed node, I didn't find any code that would ensure that these would form one connected component... so probably not?

• All these things are only conjecture based on reading this codebase. I do not make any claims that this will work or is, in any way, the thruth.