Optimizing graph layout by t-SNE perplexity estimation

Abstract

Perplexity is one of the key parameters of dimensionality reduction algorithm of t-distributed stochastic neighbor embedding (t-SNE). In this paper, we investigated the relationship of t-SNE perplexity and graph layout evaluation metrics including graph stress, preserved neighborhood information and visual inspection. As we found that a small perplexity is correlated with a relative higher normalized stress while preserving neighborhood information with a higher precision but less global structure information, we proposed our method to estimate appropriate perplexity either based on a modified standard t-SNE or the sklearn Barnes–Hut TSNE. Experimental results demonstrate effectiveness and ease of use of our approach when tested on a set of benchmark datasets.