Stochastic cluster embedding

Abstract Neighbor embedding (NE) aims to preserve pairwise similarities between data items and has been shown yield an effective principle for visualization. However, even the best existing NE methods such as stochastic neighbor (SNE) may leave large-scale patterns hidden, example clusters, despite strong signals being present in data. To address this, we propose a new cluster visualization method based on Embedding principle. We first family of that generalizes SNE by using non-normalized Kullback–Leibler divergence with scale parameter. In this family, much better visualizations often appear parameter value different from one corresponding SNE. also develop efficient software employs asynchronous block coordinate descent optimize objective functions. Our experimental results demonstrate consistently substantially improves clusters compared state-of-the-art approaches. The code our is publicly available at https://github.com/rozyangno/sce .