PolynomialInteraction

class superglm.PolynomialInteraction(poly1_name: str, poly2_name: str)

Bases: object

Cross-product of two polynomial bases.

Single group of n1 * n2 columns formed by all pairwise products of the two parents’ orthonormal components (constant excluded). Both margins evaluate through the parents’ stored factors (deep copies of the fitted parent specs) — never re-orthogonalized.

Note the parents’ weight-orthonormality does not transfer to this block: pairwise products of two weight-orthonormal margins are not weight-orthonormal unless the margins are independent under the weights, so the standardized-group-lasso argument is a main-effect property and does not extend to the product columns.

reconstruct(

beta: NDArray,

n_points: int = 50,

) → dict[str, Any]

Evaluate the surface on a grid without materialising a cross design.

The cross design over a meshgrid is (n_points**2, n1*n2), so it grows with the SQUARE of the requested resolution while carrying no information the two marginal bases do not: the product structure means f(x1_i, x2_j) = P1[i] C P2[j], which factors. At n_points=250 with two degree-50 parents the dense form is about 1.25 GB against 4.8 MB for the fit’s own design – and the rating-table impact sweep reconstructs at every rung of its ladder, so a model that exported before could run out of memory once interactions joined the sweep.

Same value and same convention as before – surface[j, i] = f(x1[i], x2[j]), which is what np.meshgrid’s default "xy" indexing produced – reached by the association TensorInteraction already uses.