stress

function stress(m): MetricResult;

Defined in: metrics/distance.ts:47

How far the projected distances are from the originals.

The classic distance-preservation measure. Sensitive to scale: a projection that halves every distance is perfect in shape but scores badly here. Use scaleNormalizedStress if the projection’s units are arbitrary, which they are for t-SNE, UMAP and most non-linear methods.

• Needs: high-dimensional data and projection. No labels.
• Range: [0, ∞), lower is better; 0 means the distances match exactly. Around 1 means the error is as large as the distances themselves.
• Cost: O(1), from an O(N²·D) pass.

Per-point values give each point’s share of the total squared error — useful for colouring, but a share rather than a score, so they sum to 1 instead of averaging to the stress. The one exception is an exactly perfect projection: with no error to apportion the array is all zeros and sums to 0, not 1.

Parameters

Parameter	Type
m	DistanceMoments

Returns

MetricResult

See

Kruskal, Psychometrika 29 (1964) https://doi.org/10.1007/BF02289565

Example

import { analyze, stress } from "@saehrimnir/sickle";

const a = analyze(data, projection);
const s = stress(a.moments);

s.value;       // 0.0807 — raw Kruskal stress, in the data's own units
s.localKind;   // "share"
s.local[0];    // 0.0104 — this point's share of the error; the array sums to 1