umap

UMAP (Uniform Manifold Approximation and Projection) is a dimension reduction technique for visualization and general nonlinear embedding, implemented as the Python package umap-learn. It is designed to work like a scikit-learn transformer, so it can be used as a drop-in alternative to t-SNE in existing pipelines. The method is based on three assumptions: data is uniformly distributed on a Riemannian manifold, the Riemannian metric is locally constant, and the manifold is locally connected. These assumptions are used to build a fuzzy topological representation, and the embedding is found by searching for a low-dimensional projection with the closest equivalent structure. The mathematics is described in the ArXiv paper 1802.03426, with a broader introduction in a Nature Reviews Methods Primers article. Installation is available through conda-forge (conda install -c conda-forge umap-learn) or PyPI (pip install umap-learn). It depends on scikit-learn, numpy, scipy, numba, tqdm, and pynndescent. Optional extras cover plotting (matplotlib, datashader, holoviews), Parametric UMAP (TensorFlow), and tbb CPU optimizations on x86. Basic usage mirrors scikit-learn: import umap from sklearn.datasets import load_digits digits = load_digits() embedding = umap.UMAP().fit_transform(digits.data) Key parameters include n_neighbors, which controls how much local versus global structure is preserved (commonly 5-50, with 10-15 a sensible default), min_dist, which controls how tightly points are packed (0.001-0.5, with 0.1 a reasonable default), and metric, which selects the distance function in input space. Sparse matrix input is supported. Documented benefits include speed on large and high-dimensional datasets, scaling in embedding dimension for general preprocessing, preservation of some global structure, support for non-metric distances such as cosine and correlation, the ability to add new points to an existing embedding via transform, and supervised or semi-supervised reduction by passing labels as y. Experimental features include inverse transform, non-Euclidean embeddings such as hyperbolic embeddings, embeddings with uncertainty, and preliminary dataframe support. The package also provides densMAP, enabled with densmap=True, which augments UMAP to preserve local density information. densMAP adds parameters such as dens_frac, dens_lambda, dens_var_shift, and output_dens, and recommends larger n_neighbors values (e.g. 30). Parametric UMAP trains a neural network to learn a UMAP-based transformation, supporting faster inference on unseen data, more robust inverse transforms, autoencoder variants, and semi-supervised classification. A umap.plot subpackage offers basic and interactive plotting with hover tools and diagnostic options. For GPU acceleration, the README points to torchdr, a PyTorch-based implementation that accelerates kNN computation, affinity construction, and embedding optimization on GPU. Documentation is hosted on Read the Docs, including a FAQ. The project is 3-clause BSD licensed and welcomes contributions. The README asks users to cite the JOSS software paper, the ArXiv paper, and, where applicable, the densMAP, Parametric UMAP, and Nature Primer references.