Context

SciPy is now favoring sparse arrays (i.e. scipy.sparse.sparray and its subclasses) over sparse matrices (i.e. scipy.sparse.spmatrix and its subclasses) to enlarge the scope of matrices to $n$-dimensional data-structures since SciPy 1.8 (see scipy/scipy#14822).

Sparse matrices now subclass their associated sparse arrays (see scipy/scipy#18440).

scikit-learn has been supporting sparse matrices but now also needs to support SciPy sparse arrays.

Proposed solutions

Ordered by preference:

• Use scipy.sparse.issparse and the format attribute everywhere and not use isinstance at all
• Use isinstance on private compatibility class defined to be scipy.sparse.spmatrix or scipy.sparse.sparray conditionally on SciPy's version (i.e. use scipy.sparse.sparray if available)
• Rely on duck-typing or the class name to check for sparse arrays

cc @ivirshup