Skip to content

Fast column-wise partial Khatri-Rao product

License

Notifications You must be signed in to change notification settings

thread/wide-product

Repository files navigation

wide-product

wide-product implements a partial, column-wise Khatri-Rao product. It is fast, and works on sparse matrices.

It can be useful for engineering of cross-features for machine learning.

Definition

For a pair of scalars (~ one by one matrices), the wide product is multiplication:

wide_product ( [[a]], [[b]] ) == [[a * b]]

Where matrices are constructed by vertical stacking, the product is row-wise:

wide_product ( vstack((A, B)), vstack((C, D)) ) ==
    vstack((wide_product(A, C),
            wide_product(B, D)))

Where matrices are constructed by horizontal stacking, the product contains all the products of the subcomponents up to permutation of columns:

wide_product ( hstack((A, B)), hstack((C, D)) ) ==
    hstack((wide_product(A, C),
            wide_product(A, D),
            wide_product(B, C),
            wide_product(B, D)))

Installation

pip install wide-product

Development

To build the module:

python setup.py build

To test:

PYTHONPATH=$(echo build/lib*):. py.test

To install:

pip install .