Advanced Preprocessing

Algorithm overview

Spline features capture smooth nonlinear relationships, target encoding converts high-cardinality categories into compact numeric features, and multi-label binarization turns sets of labels into indicator matrices. These transformations are useful when ordinary scaling or one-hot encoding does not fit the data shape.

JavaScript implementation

@kanaries/ml runs all three transformations locally in browser or Node.js. SplineTransformer creates a B-spline basis per numeric feature. TargetEncoder.fitTransform uses cross-fitting to keep a row's own target out of its encoding. MultiLabelBinarizer can return a dense matrix or the library's serializable CSR representation.

Quick start

import { utils } from '@kanaries/ml';

const spline = new utils.SplineTransformer({ nKnots: 4, degree: 3 });
const curved = spline.fitTransform([[0], [1], [2], [3]]);

const encoder = new utils.TargetEncoder({ smooth: 2, cv: 3 });
const encoded = encoder.fitTransform(
  [['free'], ['pro'], ['free'], ['team'], ['pro'], ['team']],
  [2, 12, 4, 20, 10, 22],
);

const labels = new utils.MultiLabelBinarizer();
const indicators = labels.fitTransform([['search', 'ml'], ['charts'], []]);
console.log({ curved, encoded, indicators });

Detailed API reference

SplineTransformer accepts nKnots, degree, knots (uniform, quantile, or explicit per-feature knots), extrapolation, and includeBias.

TargetEncoder accepts categories, targetType (auto, continuous, or binary), smooth, cv, shuffle, and randomState. Unknown categories map to the global target mean.

MultiLabelBinarizer accepts optional ordered classes and sparseOutput. It exposes fit, transform, fitTransform, inverseTransform, and fitted classes.