Support sparse output for sandwich products

[ElizabethSantorellaQC]

There are situations where we can expect ahead of time that the output of a sandwich product will be sparse:

• If the data is dominated by one very high-dimensional categorical. For example, say there are D dense columns, and M categorical columns. The fraction of nonzeros in the resulting sandwich product will be at most (D^2 + 2 DM + M) / (D + M)^2. If M >> D this matrix will be quite sparse. This could happen in an e-commerce pricing context, if features are a categorical product ID that is very high-dimensional and a small number of scalars.
• Perhaps we have computed x.sandwich(d1) and now we want to know x.sandwich(d2). The latter will have the same sparsity pattern as the former.
• We can analytically upper-bound the number of nonzero elements in a sandwich product of categoricals using the number of rows. If the data is made up mainly of categoricals and is very wide relative to its length, its sandwich product will be fairly sparse. This would be true in, for example, the German employer-employee data set used in the AKM papers. Since a typical worker will usually not have worked for a randomly-chosen firm, the sandwich product would be very sparse.

[zmbc]

From @stanmart in Quantco/glum#543 (comment):

as a first step I would not worry about using heuristics to try to guess whether the hessian will be sparse, but just rely on the user telling glum whether it should be

Actually, what might make even more sense is to have a class like SplitMatrix that allows the matrix to be split along both rows and columns, and continue to store parts of the sandwich product as sparse and other parts as dense. I don't know enough of the math to know what downstream effects this would have, however.

In any case, it seems like the very first step would be to make the categorical-categorical operation here use a sparse matrix. This would involve the C++ implementation here:

tabmat/src/tabmat/ext/cat_split_helpers-tmpl.cpp

Lines 44 to 94 in 2428222

	<%def name="sandwich_cat_cat(type)">
	template <typename Int, typename F>
	void _sandwich_cat_cat_${type}(
	F* d,
	const Int* i_indices,
	const Int* j_indices,
	Int* rows,
	Int len_rows,
	F* res,
	Int res_n_col,
	Int res_size
	% if type == 'complex':
	, bool i_drop_first
	, bool j_drop_first
	% endif
	)
	{
	#pragma omp parallel
	{
	std::vector<F> restemp(res_size, 0.0);
	# pragma omp for
	for (Py_ssize_t k_idx = 0; k_idx < len_rows; k_idx++) {
	Int k = rows[k_idx];
	% if type == 'complex':
	Int i = i_indices[k] - i_drop_first;
	if (i < 0) {
	continue;
	}
	% else:
	Int i = i_indices[k];
	% endif
	% if type == 'complex':
	Int j = j_indices[k] - j_drop_first;
	if (j < 0) {
	continue;
	}
	% else:
	Int j = j_indices[k];
	% endif
	restemp[(Py_ssize_t) i * res_n_col + j] += d[k];
	}
	for (Py_ssize_t i = 0; i < res_size; i++) {
	# pragma omp atomic
	res[i] += restemp[i];
	}
	}
	}
	</%def>

Unfortunately I'm not very familiar with C++. Does SciPy have a C++ API for sparse matrices?

[stanmart]

Actually, what might make even more sense is to have a class like SplitMatrix that allows the matrix to be split along both rows and columns

Creating a matrix class that can be split along both dimensions is a great end goal, but seems very complex. Having the option to return a fully sparse matrix would already be super helpful, and a great first step IMO.

Does SciPy have a C++ API for sparse matrices?

I'm not sure. What should work though is that in scipy, sparse matrices are basically a collection of 1d vectors representing the data. We could construct these vectors in C++ using the numpy API, and then convert those into a sparse matrix in python using regular scipy functions. I think it should be quite performant.

First we should figure out which kind of sparse matrix format would be the best fit for a sandwich product. After that, you could look into how those matrices are represented, and try to construct the necessary vectors. Do you think this makes sense?

[zmbc]

We could construct these vectors in C++ using the numpy API

Isn't this difficult, because the length of the vectors is not known upfront? What size would we make the numpy arrays?

[stanmart]

Yeah that's a good point. So we will need to make a guess at the density of the sandwich product at some point.