Organization of new algorithm types

[rtbs-dev]

Just finding this library, and wow! I wish I had found this sooner.

Was working on some implementations for edge filtering (see here) but eventually would love to contribute using this framework.

• Where would such algorithms go w.r.t. the others? A new submodule?
• Any style guidelines to getting started? I'll admit I am much more familiar with networkx and functional style libraries like jax, pytorch, and autograd (had to write a lot of boilerplate to keep track of node properties in those for my research). Have never seen/used scipy's LinearOperator so extensively, so any resources on using those?
• Also interested in vector similarities over graphs, see here and here. I see you guys have a lot of diffusion code already, but is simply using e.g. scipy.linalg.pinv wrapped in an Algorithm ok here?

[QLutz]

Hello,

Thank you for your interest in our package!

To answer your questions:

• Such algorithms would indeed call for a new submodule, whose name can be chosen upon your pull request. What outputs do you think those algorithms should offer?

• The guidelines for contributing can be found in the Wiki. For SciPy's LinearOperator, we relied on the documentation.

• We do not use scipy.linalg.pinv as we found it was too costly. However, if you are interested in implementing any papers in particular, please do tell us so we, in turn, can tell you if we think it would make a nice addition to the package.

Looking forward to your answer!

[rtbs-dev]

Definitely the papers I've linked! Actually, the author has implementations in pandas+nx to start from, though they are essentially single scripts.

see:

• Network backboning
• Node vector distance

Both of those would make excellent submodules, IMO. While I am familiar with the algorithms and can use those scripts as a launching point, it would honestly be a little while until I adequately understand your framework to meaningfully contribute. If you decide to add them first, however, I could certainly contribute tests, or add remaining algorithms once I have one or two to follow from.

[nathandelara]

Hi, We took a closer look at the algorithms you proposed.

• If you want to get your hands on the package, the easiest algorithm to implement is probably Diffusion maps. This would go directly into sknetwork.embedding so you can rely on examples such as sknetwork.embedding.Spectral. Most of the necessary primitives seem to be already available in sknetwork.linalg. Do not hesitate to open a dedicated issue or directly a PR so that we can help you!

• Network Distances could go into sknetwork.ranking. I would need to have a closer look at the algorithm though to check whether you need Cython or not for this one.

• Edge filtering would require to implement a new submodule from scratch, something like sknetwork.edge_filtering. So, this is slightly more challenging. Maybe we could wait until you are more familiar with the package (and we are more familiar with the algorithm...) for this one. Still, it is an interesting project!

Thanks !