UMAP Dimensionality Reduction

UMAP (Uniform Manifold Approximation and Projection) is a nonlinear dimensionality reduction method that builds a nearest-neighbor graph in high dimensions and finds a low-dimensional layout preserving its local structure.

UMAP(균일 매니폴드 근사 및 투영)은 고차원에서 최근접 이웃 그래프를 만들고 그 국소 구조를 보존하는 저차원 배치를 찾는 비선형 차원축소 방법이다.

How do I visualize high-dimensional data in two dimensions?

고차원 데이터를 2차원으로 시각화하려면 어떻게 하나요?

UMAP projects many variables into two or three dimensions by preserving each point's nearest-neighbor relationships, which makes clusters and local structure visible in a scatterplot. It fits exploratory visualization of high-dimensional data such as embeddings, omics, or many survey features, and is often preferred to t-SNE for speed.

When to use UMAP Dimensionality Reduction

• Visualizing many numeric features or embeddings in a 2D scatterplot
• Exploring whether groups or clusters exist before formal clustering
• Reducing dimensions as a preprocessing step for density-based clustering

When to use something else

• Axes must be interpretable or variance explained must be reported: use principal component analysis
• Distances between distant clusters must be meaningful: use PCA or multidimensional scaling

Assumptions and requirements

• Features are scaled appropriately
• Results depend on the number of neighbors and minimum distance settings
• Distances between clusters and cluster sizes in the plot are not directly interpretable
• Stochastic, so a random seed should be fixed for reproducibility

What the analysis reports

• Two- or three-dimensional embedding coordinates for each case
• Scatterplot colored by group or cluster
• Hyperparameters used (n_neighbors, min_dist, metric)
• Optional trustworthiness score for neighborhood preservation

How to report it

A UMAP projection (n_neighbors = [value], min_dist = [value]) of [number] features showed [description of structure].

Related methods

• t-SNE
• Principal Component Analysis (PCA)
• Multidimensional Scaling

Run UMAP Dimensionality Reduction on Quorant