@kanaries/ml
API Reference/Gaussian Processes

Gaussian Process Classification

Estimate how likely each class is on small, dense datasets using a Gaussian process classifier with an RBF kernel in browser or Node.js applications.

View as Markdown

Gaussian process classification

A Gaussian process classifier uses a kernel to express how class membership varies between nearby observations. For example, if measurements on either side of a sensor threshold are noisy, the model can return the probability of each operating state at a new measurement. With an RBF kernel, points far from all training observations approach equal class probabilities.

GaussianProcess.GaussianProcessClassifier trains on finite numeric features and labels. It approximates a binary logistic posterior using Laplace's method and supports multiple classes through one binary model per class. Each Newton update factors a dense matrix with one row and column per training sample. This makes the implementation practical for small datasets in the browser or Node.js.

Quick start

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

const state = new GaussianProcess.GaussianProcessClassifier({ lengthScale: 0.8 });
state.fit([[-2], [-1], [1], [2]], [10, 10, 20, 20]);

const classes = state.getClasses(); // [10, 20]
const [probabilities] = state.predictProba([[1.5]]);
const [label] = state.predict([[1.5]]); // 20

console.log(classes.map((value, i) => [value, probabilities[i]]));

Constructor options

OptionDefaultMeaning
lengthScale1Positive RBF length scale. Supply a number for all features or an array containing one scale per feature. A larger scale makes distant training observations more similar.
signalVariance1Positive prior variance of the latent function. The kernel equals this value when comparing a sample with itself.
maxIterPredict100Maximum Newton updates during fitting for each binary posterior.
tol1e-8Positive relative tolerance for convergence of latent function values.

The supplied kernel parameters stay fixed during training. fit(X, y) requires at least two distinct class labels and throws if a binary posterior fails to converge.

Predictions

  • predict(X) returns one numeric class label per sample.
  • predictProba(X) returns one row per sample. Columns follow the ascending numeric order from getClasses(). For multiple classes, each one-vs-rest probability is divided by the sum across classes.
  • score(X, y) returns the fraction of correctly classified samples.