umap-learn

To reduce high-dimensional data for visualization, apply UMAP to project your dataset into 2D or 3D embeddings. It supports fast, scalable transformations across domains like biology, text, and images.

Density-preserving dimensionality reduction retains the local density of original data in low-dimensional embeddings. DensMAP is the supported variant that provides density-aware representations for accurate analysis.

You can use dimensionality reduction for clustering preprocessing by applying UMAP to your high-dimensional data before running HDBSCAN clustering. This workflow helps identify meaningful group structures across domains.

Parametric UMAP supports transform for new data. It learns a parametric embedding function using neural networks, allowing you to project unseen data points into the existing low-dimensional space.

The best way to handle supervised dimensionality reduction is using supervised UMAP variants. By providing target labels during the fitting process, UMAP produces embeddings that maximize class separation for analysis.

You should not use standard UMAP when global distance preservation is critical, as it prioritizes local structure. Use DensMAP instead if preserving the density distribution of your high-dimensional data is required.