umaplearn

umaplearn, commonly referenced as umap-learn, is an open-source Python library that implements the Uniform Manifold Approximation and Projection (UMAP) algorithm for nonlinear dimensionality reduction. UMAP aims to embed high-dimensional data into a lower-dimensional space while preserving the local structure of the data and, to a lesser extent, global structure. The library is widely used for data visualization, exploratory data analysis, and as a preprocessing step for machine learning tasks.

UMAP in umap-learn provides a reusable estimator with a scikit-learn compatible API. It supports unsupervised embedding as well as supervised embedding by incorporating label information. It can produce 2D or 3D embeddings by default, with tunable parameters such as n_components, n_neighbors, and min_dist to control local detail and the balance between local and global structure. It uses the pynndescent library to compute approximate nearest neighbors for scalability on large datasets. It accepts dense arrays and sparse input, supports a variety of distance metrics, and offers options for initialization and optimization. It also exposes fit, transform, and fit_transform methods and can be integrated into pipelines.

Typical use cases include visualizing high-dimensional feature spaces in biology (e.g., single-cell data), text embeddings, image features, and other domains; it is commonly used after feature extraction with PCA or autoencoders. The library emphasizes speed and scalability and provides parallelization options and sensible defaults to balance quality and performance.

umap-learn is maintained as an open-source project with releases on PyPI and GitHub. It is distributed under the BSD 3-Clause license and has broad community support and documentation. The project continues to evolve with improvements to accuracy, speed, and usability.