Closed pauleve closed 1 year ago
Sounds like a good fit :) At the moment, we mostly considered "real world" networks but we definitely want to include random ones as well (we also have some random benchmarks to include :D). There are just two things that need to be done on my part:
I have to check if the metadata we compute (for example here) can scale to models of that size. It shouldn't be a problem, but there is a very naive SCC-decomposition algorithm for the regulatory graph used there, so that one I probably need to improve for 100K+ variable networks.
I'll let you know about this next week :)
Returning to this issue... for now, I decided to avoid including randomly generated networks. This is mostly because it is not always clear how representative they are of the "real world" cases. Also, the parameters for the network generation are still a bit unclear to me (How many networks should be included if the supply is unlimited? What is the upper/lower bound on network size? What parameters for the generator do we prefer?)
However, I do see the benefit of having some collection of larger random networks for scalability and compliance testing. For now, I have included a reference to your dataset in the project README for anyone also interested in random networks.
For benchmarking the computation of attractors with M'ost Permissive updating mode with mpbn [1], I created a set of (random) very large BNs, i.e., 1,000 to 100,000 nodes, with in-degrees ranging up to 1400
The generated models are available at https://zenodo.org/record/3714876 and stored in textual bnet format.
[1] https://nbviewer.jupyter.org/urls/zenodo.org/record/3936123/files/Scalability%20on%20large%20random%20BNs.ipynb
Let me know if I should work on a PR, or if this is out of the scope of this repository :smiley: