Statistics Glossary → Perplexity (t-SNE effective number of neighbours)
Perplexity (t-SNE effective number of neighbours)concept
The t-SNE parameter that sets the width of each point's neighbourhood and roughly indicates how many effective neighbours each point takes into account.
t-SNE assigns each point similarity probabilities to the points around it using a Gaussian kernel. The kernel width σ_i is chosen separately for each point: it is tuned by binary search so that the perplexity of this probability distribution, i.e. 2 raised to its entropy, equals the value given by the user. A narrow neighbourhood is thus used in dense regions and a wide one in sparse regions, and perplexity can be read as 'the effective number of neighbours'.
Typical values lie between 5 and 50, and the value should be clearly smaller than the number of observations. A small perplexity focuses on very local detail and can produce small clumps that do not really exist; a large perplexity preserves broader structure but smooths out local detail. Because the results are sensitive to perplexity, several values should be tried. In a t-SNE map the sizes of clusters and the distances between them are affected by perplexity and by the algorithm, and should not be interpreted directly.
An everyday example
Imagine describing the social structure of a neighbourhood by asking each person 'how many of my nearest people should I pay attention to?'. If everyone considers only 3 neighbours, the neighbourhood looks split into many small groups; if everyone considers 50, larger and coarser areas appear. Perplexity corresponds to this 'number of neighbours taken into account'.
In MerQur
The 'Perplexity' parameter of t-SNE: the effective number of neighbours of each point; with the Barnes-Hut approximation, only about the nearest 3 × perplexity neighbours are taken into account.
Analyses where this concept appears
t-SNE (t-Distributed Stochastic Neighbor Embedding)