Text Feature Extraction

Module overview

Text estimators need numeric features, but materializing a dense document-by-vocabulary matrix wastes memory. The FeatureExtraction module tokenizes strings and returns CSRMatrix sparse matrices that can flow through Pipeline and sparse-aware Naive Bayes estimators.

JavaScript implementation

@kanaries/ml brings the familiar scikit-learn text workflow to browser and Node.js code: use CountVectorizer for token counts, TfidfTransformer for weighting existing count matrices, or TfidfVectorizer for both steps together.

Quick start

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

const model = new Pipeline({ steps: [
  ['tfidf', new FeatureExtraction.TfidfVectorizer()],
  ['classifier', new Bayes.MultinomialNB()],
] });

model.fit(['red apple sweet', 'blue ocean deep'], [0, 1]);
model.predict(['sweet apple']);

Choose an API

• CountVectorizer: sparse token and n-gram counts.
• TfidfTransformer: TF-IDF weighting for an existing count matrix.
• TfidfVectorizer: one-step text-to-TF-IDF conversion.
• Hashing and Dictionary Features: bounded-memory hashing and explicit dictionary expansion.