s#

superglm.s(
column: str,
*,
kind: str = 'ps',
k: int | None = None,
n_knots: int | None = None,
degree: int = 3,
knot_strategy: str = 'uniform',
penalty: str = 'ssp',
select: bool = False,
knots: ArrayLike | None = None,
discrete: bool | None = None,
n_bins: int | None = None,
extrapolation: str = 'clip',
boundary: tuple[float, float] | None = None,
knot_alpha: float = 0.2,
constraint: ConstraintSpec | None = None,
m: int | tuple[int, ...] = 2,
lambda_policy: LambdaPolicy | dict[str, LambdaPolicy] | None = None,
) → BoundTerm#

Describe a smooth effect of one numeric column.

For example, s("age", kind="cr", k=10) declares a cubic regression spline inside a family predictor. The model learns its basis and coefficients when fitted. All spline options have the same meaning as in Spline.

Parameters:
columnstr

Name of the numeric input column.

kindstr, default=”ps”

Spline basis. Common choices are "ps" for P-splines and "cr" for cubic regression splines. See Spline for the other bases.

kint, optional

Public basis size. This limits flexibility; it is not the fitted effective degrees of freedom. Supply either k or n_knots.

n_knotsint, optional

Number of interior knots, as an alternative to k.

degreeint, default=3

Polynomial degree for bases that support this option. Cubic regression splines remain cubic.

knot_strategystr, default=”uniform”

Knot-placement rule. "uniform" spaces knots evenly; "quantile_rows" follows the training data and "quantile_tempered" tempers that concentration.

penalty{“ssp”, “none”}, default=”ssp”

Enable SSP basis reparametrization, or disable it with "none".

selectbool, default=False

Add a penalty on the spline’s null space so that smoothing can also shrink directions left unpenalized by the ordinary wiggle penalty.

knotsarray-like, optional

Explicit interior knot positions, replacing automatic placement.

discretebool, optional

Request discrete evaluation for this term. None inherits the model setting.

n_binsint, optional

Bin count for discrete evaluation. None inherits the model setting.

extrapolation{“clip”, “extend”, “error”}, default=”clip”

Prediction outside the fitted boundaries: hold the boundary value, continue the basis, or raise an error.

boundarytuple of float, optional

Explicit lower and upper spline boundaries. Otherwise use the training range.

knot_alphafloat, default=0.2

Tempering control for "quantile_tempered" knot placement.

constraintConstraintSpec, optional

Requested shape constraint. See the SuperLSS limitation below.

mint or tuple of int, default=2

Penalty order: a difference order for P-splines or a derivative order for derivative-penalty bases. Multiple orders require a basis that supports multiple penalty components.

lambda_policyLambdaPolicy or dict of str to LambdaPolicy, optional

Control whether smoothing penalties are estimated or fixed. A mapping sets policies for the spline’s individual penalty components.

Returns:
BoundTerm

A spline declaration for the named column.

See also

Spline

Basis, knot, penalty and extrapolation options.

ti

An interaction between two declared spline terms.

Notes

SuperLSS currently warns and fits unconstrained when a term requests shape constraints. A constraint argument does not enforce them there.

Examples

>>> from superglm import s
>>> age = s("age", kind="cr", k=10)
>>> age.column
'age'
On this page