umap-learn

Dimensionality reduction with UMAP compresses high-dimensional data into 2D or 3D embeddings, preserving both local and global structure for visualization. It supports sklearn-style pipelines to prepare features for downstream models.

UMAP embeddings work seamlessly for clustering with HDBSCAN. You reduce high-dimensional data to a lower-dimensional space first, then apply HDBSCAN to identify distinct clusters within the transformed embedding.

Transforming new data into an existing embedding space uses UMAP's transform options. After fitting UMAP on initial training data, you project new observations into the same learned lower-dimensional manifold.

UMAP supports supervised and semi-supervised dimensionality reduction workflows, allowing labels to guide the embedding. It also includes Parametric UMAP and alignment variants for comparing related datasets.

UMAP parameters like n_neighbors, min_dist, n_components, and metric control the embedding balance. Adjust n_neighbors for local versus global structure, and min_dist to regulate how tightly points cluster together.

UMAP may not suit purely linear dimensionality reduction needs where exact variance preservation is critical. For strict variance maximization, consider category-level linear alternatives instead of UMAP's nonlinear manifold approach.