TensorInteraction

class superglm.TensorInteraction(

feat1_name: str,

feat2_name: str,

*,

n_knots: tuple[int, int] | None = None,

decompose: bool = False,

)

Bases: object

Interaction-only tensor product spline (ti) term.

Builds centered marginal bases from parent spline specs, inheriting their knot vectors, penalties, and boundary constraints. Forms the row-wise Kronecker product to yield an interaction-only surface: constant and main-effect directions are excluded structurally.

The tensor penalty is kron(S1, I) + kron(I, S2) on the centered marginals. This leaves the bilinear x1 * x2 direction in the tensor null space while excluding the constant, x1 and x2 lower-order pieces. Group lasso can still zero the whole interaction block cleanly.

Parameters:

feat1_name, feat2_namestr

Names of the parent spline features.

n_knotstuple of int or None

(n_knots1, n_knots2) interior knots for each marginal basis. When None (default), the parent’s knot count is used directly.

decomposebool

If True, split the centered tensor basis into a 1D bilinear subgroup and a wiggly subgroup. This is useful when you want the bilinear null space to be selectable/shrinkable separately from the higher-order interaction surface.

build_discrete(

x1: NDArray,

x2: NDArray,

parent_specs: dict,

n_bins: tuple[int, int],

sample_weight: NDArray | None = None,

) → DiscreteTensorBuildResult

Build a discretized tensor basis on observed joint support pairs.

score(

x1: NDArray,

x2: NDArray,

beta: NDArray,

) → NDArray

Score the tensor interaction without materialising the row-Kronecker block.