Trustworthiness

When you reduce high-dimensional data down to 2 or 3 dimensions, the big question is: did nearby points stay nearby, or did the method scramble the local structure? Trustworthiness is a metric built to answer exactly that. It tells you how much you can trust a low-dimensional embedding to preserve the original neighborhood relationships.

What it measures

Trustworthiness focuses on false neighbors: points that look close together in the embedding but were not actually close in the original space. A dimensionality reduction method like t-SNE, UMAP, or Isomap can create visually appealing clusters, but trustworthiness checks whether those local groupings are genuine. The score ranges from 0 to 1, where values closer to 1 mean the embedding preserves local neighborhoods well. It works by comparing each point’s k nearest neighbors in the original data with its neighbors in the reduced space, then penalizing intruders that appear artificially close after projection.

Why it matters

In unsupervised learning, people often inspect embeddings to make decisions about cluster structure, anomaly patterns, or customer segments. If trustworthiness is low, the picture can be misleading:

separate customer groups can look merged,
normal transactions can appear suspiciously close to fraud cases,
document topics can seem cleaner than they really are.

This makes trustworthiness especially useful when comparing embeddings or tuning hyperparameters such as the number of neighbors in UMAP or perplexity in t-SNE.

In practice

A high trustworthiness score does not mean every global distance is preserved; it is specifically about local structure. That is why it is commonly paired with metrics like continuity or visual inspection. In Python, a common implementation is sklearn.manifold.trustworthiness, which lets you score an embedding directly. For example, if a 2D UMAP projection of product embeddings has trustworthiness 0.98, you can be fairly confident that nearby products in the plot were also nearby in the original feature space.

Trustworthiness is a metric for evaluating dimensionality reduction that measures how well local neighborhoods in a low-dimensional embedding reflect true neighborhoods in the original high-dimensional space. It penalizes points that appear as false nearest neighbors after projection, with higher values indicating better local structure preservation. Trustworthiness matters because many unsupervised analyses and visual interpretations depend on embeddings that do not invent misleading local relationships.

Imagine shrinking a detailed city map into a simple sketch. A good sketch should still keep nearby streets and landmarks near each other, even if it leaves out lots of detail. Trustworthiness measures that same idea in AI when data is squeezed into fewer dimensions for easier viewing or analysis.

It asks: when two things look close together in the simplified version, were they actually close in the original data too? If not, the picture may be misleading. So trustworthiness tells you how much you can rely on a low-dimensional view, especially for preserving local neighborhoods — the small groups of similar items that matter most.