TfidfVectorizer

Algorithm overview

TfidfVectorizer combines vocabulary learning, count vectorization, inverse-document-frequency weighting, and row normalization. Use it for document classification, similarity, and lightweight semantic search baselines.

JavaScript implementation

The @kanaries/ml implementation takes raw string[], returns a CSRMatrix, works in Pipeline, and feeds sparse-aware MultinomialNB without converting the entire corpus to dense arrays.

Quick start example

import { Bayes, FeatureExtraction, Pipeline } from '@kanaries/ml';

const pipeline = new Pipeline({ steps: [
  ['tfidf', new FeatureExtraction.TfidfVectorizer({ ngramRange: [1, 2] })],
  ['nb', new Bayes.MultinomialNB()],
] });
pipeline.fit(['red apple sweet', 'blue sea deep'], [0, 1]);
console.log(pipeline.predict(['deep blue ocean']));

The equivalent Python workflow uses Pipeline([('tfidf', TfidfVectorizer()), ('nb', MultinomialNB())]).

Detailed API reference

The constructor accepts every CountVectorizer option plus every TfidfTransformer option.

• fit(documents: string[]): void
• transform(documents: string[]): CSRMatrix
• fitTransform(documents: string[]): CSRMatrix
• getFeatureNamesOut(): string[]
• vocabulary: ReadonlyMap<string, number>
• idf: number[]

Serialize a fitted instance with JSON.stringify(vectorizer) and restore it with the root loadModel export.