snc

function snc(
   hdIn,
   ldIn,
   opts?): Snc;

Defined in: passes/snc.ts:180

Does the projection show the right groups?

Judges clusters rather than points, which is where scatterplots actually mislead: one that splits a real group in two, or fuses two into one, can still score well on trustworthiness because no single neighbourhood is badly wrong.

• Steadiness — are the groups you see real? Low means the projection shows groups that are not in the data.

• Cohesiveness — are the data’s groups still together? Low means the projection hides groups that are there.

• Needs: high-dimensional data and projection. No labels.

• Range: both [0, 1], higher is better.

• Cost: O(N²) time and O(N²) memory; maxPoints defaults to 6000.

Stochastic: clusters are drawn by random walks, so the result is an estimate. seed makes a run reproducible and more iterations narrows the spread (about ±0.005 at the default 150). Treat small differences as noise.

Parameters

Parameter	Type
hdIn	PointsInput
ldIn	PointsInput
opts	SncOptions

Returns

Snc

See

Jeon, Ko, Jo, Yi & Seo, IEEE TVCG 28 (2022) https://doi.org/10.1109/TVCG.2021.3114833

Example

import { snc } from "@saehrimnir/sickle";

// Stochastic: it draws clusters by a random walk, so fix `seed` to reproduce.
const s = snc(data, projection, { iterations: 150, seed: 1212 });

s.steadiness;     // 0.9216 — are drawn groups real?
s.cohesiveness;   // are real groups drawn together?

// Per-point contributions are opt-in:
const withLocal = snc(data, projection, { local: true });
withLocal.localSteadiness?.[0];