Model Inspection

Algorithm overview

Permutation importance measures how much a fitted model's score degrades when one feature is shuffled. Partial dependence averages predictions while forcing one or two features across a grid. Together they help identify influential inputs and visualize broad model response patterns.

JavaScript implementation

@kanaries/ml exposes model-agnostic inspection functions that work with fitted estimators in browser and Node.js. They do not depend on tree internals or linear coefficients, so the same workflow works across compatible models.

Quick start

import { Linear, utils } from '@kanaries/ml';

const X = [[0, 0], [1, 0], [2, 1], [3, 1]];
const y = [0, 2, 4, 6];
const model = new Linear.LinearRegression();
model.fit(X, y);

const importance = utils.permutationImportance(model, X, y, {
  nRepeats: 10,
  randomState: 42,
});
const dependence = utils.partialDependence(model, X, [0], {
  gridResolution: 50,
});
console.log({ importance, dependence });

Detailed API reference

permutationImportance(estimator, X, y, { nRepeats?, randomState?, scoring? }) returns importances, importancesMean, and importancesStd by feature.

partialDependence(estimator, X, features, { gridResolution?, percentiles? }) supports one or two numeric feature indices and returns gridValues plus averaged predictions shaped by the requested grid.