Statistics Glossary → Trustworthiness and continuity measures
Trustworthiness and continuity measuresconcept
Two rank-based measures of how well a low-dimensional embedding from methods such as t-SNE and UMAP preserves the local neighbourhoods of the original space.
Proposed by Venna and Kaski, these measures compare each point's k nearest neighbours in the original space and in the embedding. Trustworthiness penalizes points that appear as neighbours in the embedding but are not neighbours in the original space ('false neighbours'): a high value means that points shown side by side on the map really are similar. Continuity is the reverse: it penalizes points that are neighbours in the original space but drift apart in the embedding. The penalty is proportional to how far down the neighbour ranking the point falls; both measures range from 0 to 1, with 1 meaning perfect preservation.
They assess only local structure (the chosen k-neighbourhood); they do not say whether distances between clusters or cluster sizes are rendered faithfully. On t-SNE and UMAP maps the gaps between clusters are not suitable for interpretation in any case. The values depend on k, so different methods should be compared with the same k.
An everyday example
A music platform places 10 000 songs on a two-dimensional map by their audio features. A trustworthiness of 0.95 means that songs placed side by side really do sound alike; a continuity of 0.80 means that some songs that are actually similar ended up apart on the map, so a user looking for similar songs will miss some of them in the immediate neighbourhood.
In MerQur
In t-SNE and UMAP, the 'Trustworthiness / continuity' output gives Venna and Kaski's measures for a k = 5 neighbourhood, with distances in the original space measured using the selected metric.
Analyses where this concept appears
t-SNE (t-Distributed Stochastic Neighbor Embedding)UMAP (Uniform Manifold Approximation and Projection)
Related concepts
Perplexity (t-SNE effective number of neighbours)Stress, fit measures and the Shepard diagram (MDS)