CountVectorizer

Algorithm overview

CountVectorizer learns a vocabulary from documents and represents each document by token or n-gram frequency. It is a strong baseline for text classification, search features, and document analytics.

JavaScript implementation

The @kanaries/ml implementation accepts string[] and returns CSRMatrix, avoiding a dense allocation when the vocabulary is large. The fitted vectorizer can run in a browser or Node.js and can be serialized with the standard model codec.

Quick start example

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

const vectorizer = new FeatureExtraction.CountVectorizer({
  ngramRange: [1, 2],
  minDf: 1,
  stopWords: ['the'],
});
const counts = vectorizer.fitTransform(['the red fox', 'the red bird']);
console.log(vectorizer.getFeatureNamesOut());
console.log(counts.toDense());

Python uses CountVectorizer(ngram_range=(1, 2), stop_words=['the']); JavaScript uses camel-case option names and the same fit/transform workflow.

Detailed API reference

new FeatureExtraction.CountVectorizer({
  lowercase?: boolean;              // true
  stopWords?: string[];             // []
  ngramRange?: [number, number];     // [1, 1]
  minDf?: number;                    // 1; values below 1 are proportions
  maxDf?: number;                    // 1; values at or below 1 are proportions
  maxFeatures?: number;
  binary?: boolean;                  // false
  vocabulary?: Record<string, number> | Map<string, number>;
})

• fit(documents: string[]): void learns the vocabulary.
• transform(documents: string[]): CSRMatrix produces sparse counts.
• fitTransform(documents: string[]): CSRMatrix combines both operations.
• getFeatureNamesOut(): string[] returns terms in column order.
• vocabulary returns a defensive ReadonlyMap copy.

Because JavaScript has one numeric type, maxDf: 1 follows the proportional/default interpretation; use an integer greater than one for an absolute document threshold.