Supervised Stochastic Neighbor Embedding Using Contrastive Learning

Dimensionality reduction (DR) methods map high-dimensional data to a low-dimensional embedding, which enables data visualization. DR methods for visualization have played a critical role to gain insights into high-dimensional data, and the toolkit of DR methods has been rapidly growing in recent years (McInnes et al. [2020], Sainburg et al. [2021], Amid and Warmuth [2022], Wang et al. [2021]). Only equiped with a comprehensive understanding of these DR methods, can make informed decision based on data visualization from them, can substantially improve upon them. The state of the art for unsupervised DR relies on the stochastic neighbor embedding (SNE) framework (Hinton and Roweis [2002]), where t-SNE (van der Maaten and Hinton [2008], van der Maaten [2014]), UMAP (McInnes et al. [2020], Sainburg et al. [2021]) are two most popular example methods with impressive visualization performance on real-word data.