Transformers

Eight TransformerMixin estimators map raw functional data (n_obs, n_points) to preprocessed or reduced representations — cleaning, smoothing, re-representing, and projecting curves so downstream predictors receive well-conditioned numeric matrices.

The transformers are the workhorse of any functional-data Pipeline. A typical chain starts with Imputer to fill measurement gaps, follows with a smoother to reduce noise, and ends with FPCATransformer to project curves onto their leading functional principal components — producing an (n_obs, n_components) score matrix that any sklearn classifier or regressor can consume directly.

See the coverage / EXCLUDE list for the full 28-estimator wrapped set and all structural exclusions.

Estimator Reference

Estimator	sklearn mixin	fdars source	Key constructor params
FPCATransformer	TransformerMixin	fdars._native.regression.fpca	n_components=3, argvals=None
BSplineSmoother	TransformerMixin	fdars._native.smoothing.nadaraya_watson	bandwidth=None, kernel="gaussian", argvals=None
LocalPolynomialSmoother	TransformerMixin	fdars._native.smoothing.local_polynomial	bandwidth=None, degree=1, kernel="gaussian", argvals=None
BasisRepresentation	TransformerMixin	fdars._native.basis.fdata_to_basis_1d	n_basis=5, basis_type="bspline", argvals=None
Imputer	TransformerMixin	fdars._native.fdata (interpolation)	method="linear", constant_value=0.0, argvals=None
SplineInterpolator	TransformerMixin	fdars._native.represent.spline_interpolate	output_argvals=None, order=3, argvals=None
DepthTransformer	TransformerMixin	fdars._native.depth.fraiman_muniz_1d	depth_method="fraiman_muniz", scale=True, argvals=None
NormTransformer	TransformerMixin	fdars._native.fdata.norm_lp_1d	p=2.0, argvals=None

Estimator Details

FPCATransformer

Role in a Pipeline: the dimensionality-reduction hub. Maps (n_obs, n_points) → (n_obs, n_components) functional principal component scores. The fit call computes the FPC basis via SVD with sign canonicalization (the largest absolute value in each eigenvector is made positive), making fit idempotent across re-runs on the same data. The output score matrix feeds any sklearn classifier, regressor, or clustering estimator without adaptation.

from fdars.sklearn._skeletons import FPCATransformer
fpca = FPCATransformer(n_components=4)

BSplineSmoother

Per-curve B-spline / Nadaraya-Watson kernel smoother. Applies smoothing row-by-row so each observed curve is smoothed independently against its own evaluation grid. bandwidth=None activates automatic bandwidth selection via the native heuristic.

from fdars.sklearn._skeletons import BSplineSmoother
smoother = BSplineSmoother(bandwidth=0.2, kernel="gaussian")

LocalPolynomialSmoother

Per-curve local-polynomial kernel smoother. A polynomial of degree degree is fit locally at each evaluation point using a kernel-weighted neighbourhood. Useful when curves have local trend features that B-spline smoothing over-smooths.

from fdars.sklearn._skeletons import LocalPolynomialSmoother
lps = LocalPolynomialSmoother(degree=2, bandwidth=0.15)

BasisRepresentation

Projects each curve onto a B-spline basis of n_basis elements and returns the expansion coefficients as the transformed matrix. Output shape is (n_obs, n_basis). A 1-feature guard prevents the native call from receiving degenerate single-column inputs.

from fdars.sklearn._skeletons import BasisRepresentation
basis = BasisRepresentation(n_basis=8, basis_type="bspline")

Imputer

Linear-interpolation imputation of NaN values in the raw curve matrix. Each curve is imputed independently; boundary NaNs are forward/backward filled. Place Imputer first in any Pipeline when measurement gaps are expected.

from fdars.sklearn._skeletons import Imputer
imp = Imputer(method="linear")

SplineInterpolator

Resamples each curve from its current argvals grid onto a new output_argvals grid using spline interpolation of order order (default 3, cubic). Useful for harmonising irregular or mis-aligned grids before downstream estimators that expect a fixed grid.

from fdars.sklearn._skeletons import SplineInterpolator
interp = SplineInterpolator(output_argvals=np.linspace(0, 1, 50), order=3)

DepthTransformer

Maps each curve to a single scalar: its modified band-depth (or the depth function named by depth_method). Output shape is (n_obs, 1). Useful for constructing a depth-based feature column that feeds a standard scalar classifier or for downstream outlier ranking.

from fdars.sklearn._skeletons import DepthTransformer
dt = DepthTransformer(depth_method="fraiman_muniz", scale=True)

NormTransformer

Maps each curve to its L^p norm scalar. Output shape is (n_obs, 1). With p=2.0 (default) this is the functional L² norm; p=1.0 gives the L¹ norm. A lightweight feature extractor when only global curve magnitude matters.

from fdars.sklearn._skeletons import NormTransformer
nt = NormTransformer(p=2.0)

Typical Pipeline

from sklearn.pipeline import Pipeline
from fdars.sklearn._skeletons import Imputer, BSplineSmoother, FPCATransformer, FPCLDAClassifier

pipe = Pipeline([
    ("imputer",  Imputer()),
    ("smoother", BSplineSmoother()),
    ("fpca",     FPCATransformer(n_components=4)),
    ("clf",      FPCLDAClassifier()),
])

The transformer chain is order-sensitive by design — the output of each step must match the input contract of the next. See the concept overview on sklearn/index.md for the plain-ndarray contract.