Preprocessing Utilities

Algorithm overview

Preprocessing utilities convert raw arrays into model-ready features. They cover scaling, normalization, label and categorical encoding, binarization, missing-value imputation, variance filtering, and univariate feature selection.

JavaScript implementation

@kanaries/ml exports these helpers under utils.Preprocessing, so feature preparation can run in the same JavaScript or TypeScript runtime as model training and prediction.

Quick start example

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

const scaler = new utils.Preprocessing.StandardScaler();
const scaled = scaler.fitTransform([[1, 10], [2, 20], [3, 30]]);

const encoder = new utils.Preprocessing.OneHotEncoder({ drop: 'ifBinary' });
const encoded = encoder.fitTransform([
  ['red', 'S'],
  ['blue', 'M'],
]);

Detailed API reference

Scalers and normalizers

new utils.Preprocessing.StandardScaler({ withMean?: boolean; withStd?: boolean })
new utils.Preprocessing.MinMaxScaler({ featureRange?: [number, number] })
new utils.Preprocessing.MaxAbsScaler()
new utils.Preprocessing.Normalizer({ norm?: 'l1' | 'l2' | 'max' })

Scaler methods:

• fit(X: number[][]): void
• transform(X: number[][]): number[][]
• fitTransform(X: number[][]): number[][]
• inverseTransform(X: number[][]): number[][] for StandardScaler, MinMaxScaler, and MaxAbsScaler

Normalizer supports fit, transform, and fitTransform; fit is a no-op.

Encoders and binarizer

new utils.Preprocessing.LabelEncoder()
new utils.Preprocessing.OrdinalEncoder()
new utils.Preprocessing.OneHotEncoder({ drop?: 'none' | 'first' | 'ifBinary' })
new utils.Preprocessing.Binarizer({ threshold?: number })

LabelEncoder works on numeric label arrays. OrdinalEncoder and OneHotEncoder accept categorical matrices containing string, number, boolean, or null values.

Imputation and feature selection

new utils.Preprocessing.SimpleImputer({
  strategy?: 'mean' | 'median' | 'mostFrequent' | 'constant';
  fillValue?: number;
  missingValues?: number | null;
})

new utils.Preprocessing.VarianceThreshold({ threshold?: number })
utils.Preprocessing.fRegression(X: number[][], y: number[]): number[]
new utils.Preprocessing.SelectKBest({ k?: number; scoreFunc?: FeatureScoreFunc })

SimpleImputer supports fit, transform, and fitTransform. VarianceThreshold and SelectKBest remove features during transform.

JavaScript deployment notes

• Fit preprocessing on training data only, then reuse the fitted transformer for inference.
• Keep encoder category assumptions stable between browser sessions or Node.js jobs.
• Use fitTransform for quick experiments, but keep explicit fit and transform calls in production pipelines.