umap-learn

Dimensionality reduction transforms high-dimensional data into low-dimensional embeddings while preserving local and global structure. UMAP performs this nonlinear reduction efficiently, standardize your features first, fit the embedding, then visualize the resulting 2D or 3D coordinates.

Yes. UMAP reduces dimensionality to prepare data for density-based clustering algorithms like HDBSCAN. The low-dimensional embeddings make cluster detection more reliable by compressing noise and separating dense regions.

UMAP supports both supervised and semi-supervised modes, allowing you to incorporate label information when available. This produces embeddings that reflect both data structure and class relationships for more interpretable visualizations.

Standardize your numerical features as a critical preprocessing step before fitting UMAP. This ensures all dimensions contribute equally to the distance calculations and improves embedding quality.

Use parametric UMAP to create a model that transforms new data consistently. After fitting on your training set, apply the same model to new samples to generate embeddings aligned with the original embedding space.

UMAP supports multiple distance metrics including Euclidean, Manhattan, and others. You can tune the metric parameter alongside n_neighbors, min_dist, and n_components to customize the embedding for your data and use case.