Visible to the public Trustworthiness of t-Distributed Stochastic Neighbour Embedding

TitleTrustworthiness of t-Distributed Stochastic Neighbour Embedding
Publication TypeConference Paper
Year of Publication2016
AuthorsPandey, Shishir, Vaze, Rahul
Conference NameProceedings of the 3rd IKDD Conference on Data Science, 2016
PublisherACM
Conference LocationNew York, NY, USA
ISBN Number978-1-4503-4217-9
Keywordscomposability, dimension reduction, nonlinear embedding, pubcrawl, t-SNE, trustworthiness, trustworthy
Abstract

A well known technique for embedding high dimensional objects in two or three dimensional space is the t-distributed stochastic neighbour embedding (t-SNE). The t-SNE minimizes the Kullback-Liebler (KL) divergence between two probability distributions, one induced on points in the high dimensional space and the other induced on points in the low dimensional embedding space. In this work, we consider a more general framework of using Renyi divergence which is parametrized by the order a, the KL-divergence is a special case when a - 1.We study how various Renyi divergences perform when compared to the KL-divergence. We show that in terms of the metrics of trustworthiness and neighbourhood preservation, the embedding becomes better as Renyi divergence approaches the KL-divergence.

URLhttp://doi.acm.org/10.1145/2888451.2888465
DOI10.1145/2888451.2888465
Citation Keypandey_trustworthiness_2016