Kernel methods

Algorithm overview

Kernel methods compare samples through a similarity function instead of explicitly expanding nonlinear features. KernelRidge solves regularized regression in the dual; KernelDensity estimates a probability density from nearby observations.

JavaScript implementation

@kanaries/ml keeps both estimators in the JavaScript runtime: fit nonlinear regression in a Node.js service, or score small density models directly in an interactive browser workflow without a Python inference endpoint. Choose this implementation when model state and product code need to share the same typed API; benchmark larger training sets because kernel matrices and pairwise scoring grow with the number of samples.

Interactive Kernel Ridge regression playground

Tune the RBF bandwidth through gamma, switch datasets, and add your own observations. The nonlinear curve is fitted live with Kernel.KernelRidge.

Quick start

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

const X = [[0], [1], [2]];
const regressor = new Kernel.KernelRidge({ kernel: 'rbf', gamma: 0.5 });
regressor.fit(X, [0, 1, 4]);
const density = new Kernel.KernelDensity({ bandwidth: 0.8 });
density.fit(X);
console.log({ prediction: regressor.predict([[1.5]]), logDensity: density.scoreSamples([[1.5]]) });

Detailed API reference

KernelRidge supports linear, polynomial, RBF, sigmoid, and cosine kernels. KernelDensity supports Gaussian, tophat, Epanechnikov, exponential, linear, and cosine radial kernels and exposes scoreSamples and score.