Univariate feature scores

Algorithm overview

Univariate scores examine one feature at a time. Chi-square is suited to non-negative count features, ANOVA F measures class separation, and mutual information detects more general nonlinear dependence using discrete counts or k-nearest-neighbor entropy estimates.

JavaScript implementation

@kanaries/ml makes these score functions available in browser and Node.js code through FeatureSelection, including seeded jitter for repeated continuous values.

Quick start example

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

const X = [[0, 1], [1, 0], [5, 1], [6, 0]];
const y = [0, 0, 1, 1];
const [fScores, pValues] = FeatureSelection.fClassif(X, y);
const information = FeatureSelection.mutualInfoClassif(X, y, { randomState: 42 });
console.log({ fScores, pValues, information });

Detailed API reference

• chi2(X, y) and fClassif(X, y) return [scores, pValues].
• mutualInfoClassif(X, y, options?) and mutualInfoRegression(X, y, options?) return one non-negative score per feature.
• Mutual-information options are discreteFeatures?: boolean | boolean[], nNeighbors?: number (default 3), and randomState?: number.