Overview
Normalizer scales individual samples to have unit norm. This is different from other scalers which operate on features (columns). Each sample (row) is normalized independently.
Constructor
new Normalizer(options?: NormalizerOptions)
options.norm
'l1' | 'l2' | 'max'
default:"'l2'"
The norm to use to normalize each non-zero sample:
'l1': Sum of absolute values equals 1'l2': Sum of squares equals 1 (Euclidean norm)'max': Maximum absolute value equals 1
Properties
Number of features seen during fit.
Methods
fit
Fit the normalizer. This method only records the number of features for validation.
Returns: this - The fitted normalizer instance.
transform(X: Matrix): Matrix
Data matrix to normalize. Each row will be normalized independently.
Matrix - Normalized data where each row has unit norm.
fitTransform(X: Matrix): Matrix
fit(X).transform(X).
Training data matrix to fit and normalize.
Matrix - Normalized data.
Example
import { Normalizer } from 'bun-scikit';
const X = [
[1, 2, 3],
[4, 5, 6],
[7, 8, 9]
];
// L2 normalization (default)
const normL2 = new Normalizer({ norm: 'l2' });
const XL2 = normL2.fitTransform(X);
console.log('L2 normalized:', XL2);
// Each row has Euclidean norm of 1
// Row 0: [0.267, 0.534, 0.802]
// L1 normalization
const normL1 = new Normalizer({ norm: 'l1' });
const XL1 = normL1.fitTransform(X);
console.log('L1 normalized:', XL1);
// Each row sums to 1
// Row 0: [1/6, 2/6, 3/6] = [0.167, 0.333, 0.5]
// Max normalization
const normMax = new Normalizer({ norm: 'max' });
const XMax = normMax.fitTransform(X);
console.log('Max normalized:', XMax);
// Each row's max absolute value is 1
// Row 0: [1/3, 2/3, 1] = [0.333, 0.667, 1]
Example: Text Feature Vectors
import { Normalizer } from 'bun-scikit';
// Term frequency vectors for documents
const termFrequencies = [
[3, 0, 1], // Document 1: 3x term1, 0x term2, 1x term3
[2, 0, 0], // Document 2: 2x term1, 0x term2, 0x term3
[3, 0, 0], // Document 3: 3x term1, 0x term2, 0x term3
];
// Normalize to unit length for similarity comparison
const normalizer = new Normalizer({ norm: 'l2' });
const normalized = normalizer.fitTransform(termFrequencies);
console.log('Normalized TF vectors:', normalized);
// Now we can compute cosine similarity by dot product
Example: Zero-Norm Rows
import { Normalizer } from 'bun-scikit';
const X = [
[1, 2, 3],
[0, 0, 0], // Zero vector - cannot be normalized
[4, 5, 6]
];
const normalizer = new Normalizer();
const XNorm = normalizer.fitTransform(X);
console.log('Normalized:', XNorm);
// Row with all zeros remains unchanged
// [[0.267, 0.534, 0.802], [0, 0, 0], [0.456, 0.570, 0.684]]
Notes
- Normalization is performed row-wise (per sample), not column-wise (per feature)
- Rows with zero norm remain unchanged (all zeros)
- Unlike other scalers, the fit step doesn’t compute statistics - it only validates the data
- Commonly used in text classification and clustering
- Input data must be finite (no NaN or Infinity values)