New guidance for using t-SNE: Alternative defaults, hyperparameter selection automation, and comparative evaluation

• Published in: Visual informatics (Online) Vol. 6; no. 2; pp. 87 - 97
• Main Authors: Gove, Robert, Cadalzo, Lucas, Leiby, Nicholas, Singer, Jedediah M., Zaitzeff, Alexander
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2022; Elsevier
• Subjects: Dimensionality reduction; Machine learning; t-SNE; 99-00; Dimensionality reduction; t-SNE; 00-01; Machine learning
• ISSN: 2468-502X; 2468-502X
• DOI: 10.1016/j.visinf.2022.04.003

We present new guidelines for choosing hyperparameters for t-SNE and an evaluation comparing these guidelines to current ones. These guidelines include a proposed empirically optimum guideline derived from a t-SNE hyperparameter grid search over a large collection of data sets. We also introduce a new method to featurize data sets using graph-based metrics called scagnostics; we use these features to train a neural network that predicts optimal t-SNE hyperparameters for the respective data set. This neural network has the potential to simplify the use of t-SNE by removing guesswork about which hyperparameters will produce the best embedding. We evaluate and compare our neural network-derived and empirically optimum hyperparameters to several other t-SNE hyperparameter guidelines from the literature on 68 data sets. The hyperparameters predicted by our neural network yield embeddings with similar accuracy as the best current t-SNE guidelines. Using our empirically optimum hyperparameters is simpler than following previously published guidelines but yields more accurate embeddings, in some cases by a statistically significant margin. We find that the useful ranges for t-SNE hyperparameters are narrower and include smaller values than previously reported in the literature. Importantly, we also quantify the potential for future improvements in this area: using data from a grid search of t-SNE hyperparameters we find that an optimal selection method could improve embedding accuracy by up to two percentage points over the methods examined in this paper.