nerv

function nerv(p): MetricResult & {
  precision: number;
  recall: number;
};

Defined in: metrics/embedding.ts:148

Neighbourhood preservation framed as an information-retrieval trade-off.

Treats each point’s neighbourhood as a probability distribution and measures the divergence between the two spaces. recall is the cost of neighbours the projection missed; precision is the cost of neighbours it invented. lambda chooses which matters: 1 is pure recall, 0 pure precision, 0.5 balanced.

• Needs: high-dimensional data and projection. No labels.
• Range: [0, ∞), lower is better; 0 only when the neighbourhood distributions coincide.
• Cost: O(N²·D) for its own pass, plus a bisection per point.

Reported per point rather than summed, so values are comparable across dataset sizes.

Parameters

Parameter	Type
p	Nerv

Returns

See

Venna, Peltonen, Nybo, Aidos & Kaski, JMLR 11 (2010) https://www.jmlr.org/papers/v11/venna10a.html

Example

import { nervPass, nerv } from "@saehrimnir/sickle";

// nerv reads a NeRV pass, not `analyze`.
const p = nervPass(data, projection, { lambda: 0.5, perplexity: 30 });

nerv(p).value;  // 0.4611 — λ·recall + (1-λ)·precision, lower is better
p.recall;       // 77.6049 — KL(p‖q), the SNE/t-SNE end at λ = 1
p.precision;    // 106.8424 — KL(q‖p), the λ = 0 end