libpysal.graph roadmap to release

[martinfleis]

I tried to outline what is missing before we could cut a release with the new graph stuff. Open to a discussion.

testing

We have a decent coverage of base but especially constructors are not always tested for correctness. So we test the API but not that the adjacency captures what it should. So missing tests:

• test correctness of contiguity builders finalise contiguity graph builders #566
• test correctness of triangulation builders
• test correctness of kernel builders
• test all kernel parameters
• test all query options (both sklearn and scipy)
• test precomputed distance matrix support (indirectly via triangulation and distance band already covered)

implementation/checks

Some bits are still waiting for implementation (I have likely missed stuff here) or a double check that the current implementation works as intended.

• isolates in kernel and knn
• co-located in knn
• knn haversine checks
• fix cliques
• set operations
• draft API and raise NotImplementedError?

discussion

A few things I'd like to discuss (during the dev call in 3 minutes?)

• default kernel - W has triangular, we now have gaussian following scipy model - do we want to warn? how?
• expose kernels in transform? kernels are essentially transformations. Shall we allow to use them post-creation?
• how should set operations work? See perimeter weights, fix kernel, set_ops #561

follow-up

Stuff that is not implemented or started but that can likely wait for subsequent release, not the initial one. Below is the list of stuff I'd like to see soon but there's more to be done.

• lag_categorical (inside lag()?)
• classic I/O (we have custom Parquet now)
• plotting ENH: add vectorized plotting to Graph #593
• fuzzy contiguity implement fuzzy contiguity on Graph #564
• other utils - any priorities?
• optimal sparse formats (coo may be slow) REF: minor performance improvements in Graph #697
• IDs handling cleanup allow ids in graph when passing sparse #626
• symmetrize (see Interest in a mutual knn? #664)
• fill diagonal fill diagonal in graph #668
• higher-order neighbors
• trace

documentation

API reference should be in the docs for the first release. user guide can follow later.

• API reference
• user guide
• migration guide

Some of the follow-up could be piped via W for the time being if we want them available.

[martinfleis]

Can you drop a note if you are planning to work on any of the topics mentioned above so we don't overlap?

[ljwolf]

I am unlikely to have time until sept 27

[knaaptime]

I think I’ll go back to ensuring the constructors can consume sparse next

[martinfleis]

With #577, we're ready for the initial experimental release of Graph.

• cliques are not fixed, but they raise NotImplementedError at the moment.
• precomputed matrix is not tested explicitly but a few of our internal functions use it, so it is tested via those.
• API is not yet drafted, but that does not need to be for a release
• The default kernel is still Gaussian
• set operations are implemented with some restrictions ensuring the validity of the Graph and its alignment with original data. See implement set ops and to_networkx #575 for details.

Anything else you'd like to get in?

I would target the end of the week at the latest to cut 4.8.0 (need it alive on conda on Tuesday)

[martinfleis]

Anything blocking the release now that #577 is merged?

@jGaboardi any idea if the current infrastructure works as intended? It has not been updated here.

[jGaboardi]

I am not sure. Let's give it try; perhaps as a 4.7.0.post1 release?