Feature hashing

Algorithm overview

Hashing maps features into a fixed-width sparse matrix without storing a vocabulary. Dictionary vectorization instead learns explicit names, which is preferable when interpretability matters.

JavaScript implementation

HashingVectorizer tokenizes text, FeatureHasher accepts dictionaries, pairs, or strings, and DictVectorizer expands numeric and categorical dictionary values.

Quick start

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

const documents = ['red fox', 'blue fox'];
const X = new FeatureExtraction.HashingVectorizer({ nFeatures: 4096 }).fitTransform(documents);
console.log(X.shape);

Detailed API reference

Both hashers use sklearn-compatible MurmurHash3 indices and signed collisions. DictVectorizer exposes vocabulary, getFeatureNamesOut(), inverseTransform(), and restrict(), and can return CSR or dense output.