Isomap

Algorithm overview

Isomap builds a nearest-neighbor graph, computes all-pairs shortest-path distances, and embeds those geodesic distances with classical multidimensional scaling. It is useful when Euclidean distance cuts across a curved manifold but graph paths follow its true geometry.

JavaScript implementation

Manifold.Isomap provides both fitTransform and out-of-sample transform in browser and Node.js. The implementation stores an O(n²) distance matrix and uses an O(n³) shortest-path phase, so it is intended for moderate interactive datasets rather than unbounded production traffic.

Quick start example

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

const X = [[0, 0], [1, 0], [2, 1], [3, 1], [4, 2]];
const model = new Manifold.Isomap({ nNeighbors: 2, nComponents: 2 });
const embedding = model.fitTransform(X);
const newPoints = model.transform([[2.5, 1]]);
console.log({ embedding, newPoints });

Detailed API reference

new Manifold.Isomap({ nNeighbors?: number, nComponents?: number }) defaults to 5 neighbors and 2 components. Methods are fit(X), fitTransform(X), and transform(X). Read-only embedding and distMatrix properties expose the fitted coordinates and geodesic distances. If the neighbor graph is disconnected, the implementation warns and joins each component pair at its closest Euclidean samples, matching sklearn's non-precomputed recovery path.