umap-learn

UMAP dimensionality reduction transforms high-dimensional data into low-dimensional embeddings that preserve neighborhood structure for exploratory visualization and clustering preprocessing. It creates 2D or 3D embeddings ready for plotting and downstream modeling.

For UMAP embeddings, tune n_neighbors and min_dist differently to support clean separation for HDBSCAN-style clustering or smoother exploratory visualization. Use n_neighbors=15 and min_dist=0.1 as a starting point for standard 2D plots.

UMAP supports supervised and semi-supervised embedding by using labels (y) or partial labels (with -1) to guide embeddings for class separation. This target-driven fitting helps when target information exists for your dataset.

Standardizing features is required before running UMAP dimensionality reduction. You must standardize your dataset first, then apply UMAP with configured parameters like n_neighbors, min_dist, n_components, and metric to produce meaningful embeddings.

UMAP integrates with scikit-learn workflows by producing low-dimensional embeddings for feature engineering and downstream classifiers. It supports fit/transform operations to generate features that preserve neighborhood structure for modeling tasks.