umap-learn

Dimensionality reduction with UMAP generates fast low-dimensional embeddings that preserve meaningful manifold structure, making it ideal for visualizing high-dimensional datasets in 2D or 3D while maintaining local neighborhoods.

Prior feature standardization is required before generating UMAP embeddings. You should apply StandardScaler to your tabular numeric data or document vectors to ensure neighborhood distance parameters compute accurately during the fit_transform workflow.

Supervised UMAP incorporates label information via the target parameter (y) to improve class separation in the resulting embeddings. This supports both fully labeled and partially labeled datasets for semi-supervised learning workflows.

UMAP serves as clustering-friendly preprocessing for density-based clustering workflows like HDBSCAN. It produces embeddings optimized to preserve local manifold structure, which helps density-based algorithms identify meaningful clusters in high-dimensional data.

The n_neighbors parameter controls local neighborhood size, while min_dist controls minimum distance between embedded points. You can also configure the distance metric (e.g., euclidean) and n_components to customize the embedding output for your specific dataset.

UMAP may not be suitable when strict preservation of global distance relationships is required, as it prioritizes local neighborhood structure. Additionally, without prior feature standardization, distance calculations on unscaled high-dimensional data can produce distorted embeddings.