CSC/ CSR classes

Is your feature request related to a problem? Please describe.

Could explicit CSR and CSC classes be added to this library?

I would love to stop using scipy.sparse.spmatrix classes. Their matrix API is a source of constant frustration. Lack of CSC and CSR arrays with an np.ndarray API has been a barrier to adoption not only of this library, but of xarray and dask for the scanpy/ anndata projects

Describe the solution you'd like

I would like there to be explicit CSR and CSC classes in this package. These should perform at least as well as scipy.sparse.spmatrix counterparts.

While higher dimensional algorithms are developed, performant implementations for these cases can be taken from scipy.sparse or developed from reference implementations. I think this is a low effort/ high pay-off path towards performant implementations for common formats. They wouldn't be n-dimensional, but there's a lot of value in being able to start consolidating efforts to this library before an n-dimensional generalization has the features and performance of the common 2d cases.

Describe alternatives you've considered

No explicit CSC/ CSR classes. These are subsets of more generic structures so why special case them?

I think this is the wrong approach due to how common these structures are. Even a flexible framework like taco exports these formats (pytaco.csc and pytaco.csr).

We could also consider what code looks like if these classes are not defined. All dispatch here could change from isinstance(x, sparse.csr_matrix) to something like isinstance(x, sparse.GXCS) and x.compressed_axes == (0,) and x.ndims == 2), but something closer to the first case is probably an easier sell.

CSC/ CSR classes as thin wrappers of an n-dimensional format

I think this is a great goal. However, it requires an implementation of that format which is at least as performant in most (probably all) cases as the scipy equivalents for broad adoption. As mentioned above, performant implementations can be (1) sourced from a current dependency (scipy.sparse) or (2) easily adapted from existing reference implementations.

Code specialized for these cases could gradually be replaced with more generic versions as they are implemented. However, performant implementation for these common cases would not be blocked by implementation of n-dimensional algorithms.

Additional context

I could expand on why I think this is the right way to go

I could go on about how important I think it is for there to be CSR and CSC classes in this library, and why it's important to implement these separately from higher dimensional generalizations. I find that makes this post too long. I'd be happy to add these arguments if asked.

Type hierarchy

I don't know what the appropriate type hierarchy is here. Should issubclass(CSR, GXCS) work? Or should there be an abstract CompressedDimSparseArray parent class for all of these? What is the result type of np.stack(: list[CSR]).

These are things that could be decided on later, once the n-dimensional compressed representation is public API.

The issue does not name implementation files or tests. Start by examining the existing sparse-array type hierarchy and the scipy.sparse CSR and CSC counterparts, then determine the public API, inheritance, and performance requirements before implementing anything. Done means explicit CSR and CSC classes with ndarray-like behavior and performance comparable to the scipy equivalents.