How to Use t-SNE Effectively

By Martin Wattenberg, Fernanda B. Viégas, I. Johnson. 2016

853 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Sociology and Political Science, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Sociology and Political Science

In Representation & Self-Supervised Learning

Citations in the corpus

Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges (2021, background): “Wattenberg et al. (2016); Coenen and Pearce (2019) discuss possible pitfalls in data analysis created by t-SNE and UMAP due to loss of non-local information.”
Understanding How Dimension Reduction Tools Work: An Empirical Approach to Deciphering t-SNE, UMAP, TriMAP, and PaCMAP for Data Visualization (2020, methodology): “For instance, t-SNE can be very sensitive to the perplexity parameter, and creates spurious clusters; both t-SNE and UMAP perform beautifully in preserving local structure but struggle to preserve global structure (Wattenberg et al., 2016; Coenen and Pearce, 2019).”
Automated optimized parameters for T-distributed stochastic neighbor embedding improve visualization and analysis of large datasets (2019, methodology): “The t-SNE algorithm can be guided by a set of parameters that ﬁ nely adjust multiple aspects of the t-SNE run 19 .”
Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires (2020, methodology): “…dataset, like the distances between clusters in low-dimensional embeddings, can be less meaningful in graph-based dimensionality reduction than in PCA or MDS, because current graph-based methods tend to local-notions of distance like nearest neighbors to construct a graphical representation [33].”
Heavy-tailed kernels reveal a finer cluster structure in t-SNE visualisations (2019, background): “It is well-known that t-SNE, especially with low perplexity, can find “clusters” in pure noise, picking up random fluctuations in the density [17].”
A Survey of Imitation Learning Methods, Environments and Metrics (2024, background): “…(i) they are highly sensitive to hyperparameters, as pointed out by van der Maaten and Hinton [109]; (ii) interpreting them might be challenging [113], which makes them less suitable as an explainable method; and (iii) and some of these techniques lack robustness and might not provide…”
Imitation Learning: A Survey of Learning Methods, Environments and Metrics (2024, neutral): “…(i) they are highly sensitive to hyperparameters, as pointed out by van der Maaten and Hinton [97]; (ii) interpreting them might be challenging [98], which makes them less suitable as an explainable method; and (iii) and some of these techniques lack robustness and might not provide consistent…”
Finding Universal Grammatical Relations in Multilingual BERT (2020, background): “We acknowledge as well issues in interpreting t-SNE plots (Wattenberg et al., 2016), and include multiple plots with various hyperparameter settings to hedge against this confounder in Figure 11.”
Evaluation of UMAP as an alternative to t-SNE for single-cell data (2018, background): “As noted by others[11], t-SNE expands low density areas and tends to ignore global relationships.”
Understanding Generalization through Visualizations (2019, neutral): “ generalization. (Hochreiter and Schmidhuber, 1997) ﬁrst proposed that ﬂat minima tend to generalize well. This idea was reinvigorated by (Keskar et al., 2t-SNE analysis, following the guidelines in (Wattenberg et al., 2016). 4 (a) 100% train, 100% test (b) 100% train, 7% test (c) Minimizer of netwo”
Deep learning multidimensional projections (2019, methodology): “: During inference, our method simply executes a trained neural network, which requires no parameter setting. There is no need for guessing the ‘right’ values of parameters such as t-SNE’s perplexity [Wat16]. During training, the only free parameter to be set is the maximal loss or, alternatively, n”
Metabolomics and Multi-Omics Integration: A Survey of Computational Methods and Resources (2020, background): “, perplexity) can produce radically different plots, and misleading apparent clusters can result from random data [61].”
Efficient Algorithms for t-distributed Stochastic Neighborhood Embedding (2017, background): “In general, the location of clusters with respect to one another is meaningless, and their 2D shape is not interpretable and dependent on initialization (Wattenberg et al. (2016)).”
Automated Optimal Parameters for T-Distributed Stochastic Neighbor Embedding Improve Visualization and Allow Analysis of Large Datasets (2018, methodology): “The t-SNE algorithm can be guided by a set of parameters that finely adjust multiple aspects of the t-SNE run 21 .”
A high-bias, low-variance introduction to Machine Learning for physicists (2018, neutral): “, 2017) models, lattice gauge theories (Wetzel and Scherzer, 2017), and different phases of polymers (Wei et al., 2017).”