Self-Training Classifier

Algorithm overview

Self-training repeatedly fits a probabilistic classifier, assigns pseudo-labels to the most confident unlabeled rows, and refits with the expanded labeled set. It helps when labels are scarce but class probabilities are reasonably calibrated.

JavaScript implementation

@kanaries/ml wraps any registered classifier with predictProba, including GaussianNB. Use -1 for unlabeled targets and choose a confidence threshold or a fixed number of additions per iteration.

Quick start example

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

const model = new SemiSupervised.SelfTrainingClassifier({
  estimator: new Bayes.GaussianNB(), threshold: 0.8, maxIter: 10,
});
model.fit([[-3], [-2], [-1], [1], [2], [3]], [0, 0, -1, -1, 1, 1]);
console.log(model.transduction);

Detailed API reference

Options: estimator, threshold (default .75), criterion: 'threshold' | 'kBest', kBest (default 10), and maxIter (default 10).

Methods: fit, predict, predictProba. Learned fields: transduction, labeledIteration, nIter, and terminationCondition. Seed labels retain iteration 0; labels that remain unknown retain -1.