t-SNE Dimensionality Reduction

t-SNE (t-distributed stochastic neighbor embedding) is a nonlinear dimensionality reduction technique that maps high-dimensional data into two or three dimensions while preserving local neighborhood structure, mainly for visualization.

t-SNE(t-분포 확률적 이웃 임베딩)는 고차원 데이터를 지역적 이웃 구조를 보존하면서 2차원이나 3차원으로 사상하는 비선형 차원축소 기법으로, 주로 시각화에 쓰입니다.

How do I visualize high-dimensional data in two dimensions to see whether it forms groups?

고차원 데이터를 2차원으로 시각화해서 집단이 형성되는지 어떻게 확인하나요?

t-SNE projects high-dimensional observations into a two- or three-dimensional map in which points that are close neighbors in the original space stay close together. It fits exploratory visualization of possible groupings, but distances between clusters and cluster sizes in the map should not be interpreted quantitatively.

When to use t-SNE Dimensionality Reduction

• Visualizing many numeric features, such as embeddings or expression profiles, in 2D
• Checking visually whether known labels separate in the feature space
• Exploring possible cluster structure before formal clustering

When to use something else

• Interpretable components or variance explained are needed: use principal component analysis
• New observations must be projected or more global structure preserved: use UMAP

Assumptions and requirements

• Numeric features scaled to comparable ranges
• Perplexity chosen with care, as results depend on it
• Random seed fixed, because the embedding is stochastic
• Cluster sizes and between-cluster distances are not meaningful

What the analysis reports

• Two- or three-dimensional embedding coordinates
• Scatter plot colored by a grouping variable
• Final Kullback-Leibler divergence
• Settings used (perplexity, iterations, learning rate)

How to report it

A t-SNE projection (perplexity = [value]) of the [n] observations showed [description of pattern] when points were colored by [variable].

Related methods

• Principal Component Analysis (PCA)
• UMAP
• Multidimensional Scaling

Run t-SNE Dimensionality Reduction on Quorant