umap-learn

UMAP provides fast dimensionality reduction to generate low-dimensional embeddings from high-dimensional data for visualization. You train a model on your dataset to obtain 2D coordinates that preserve local manifold structure.

UMAP serves as clustering preprocessing by reducing data to a low-dimensional manifold representation. The resulting embeddings can be fed into clustering algorithms to group similar data points more effectively.

Yes, UMAP supports parametric, DensMAP, and supervised variants. Parametric UMAP uses neural networks for embeddings, DensMAP preserves local density, and supervised mode incorporates labels for task-specific dimensionality reduction.

Yes, UMAP provides transform and inverse_transform functions. Transform projects new data into the learned embedding space, while inverse_transform maps low-dimensional coordinates back to the original high-dimensional space.

UMAP offers an aligned UMAP feature to find coherent low-dimensional embeddings across related datasets. This ensures that shared structures remain consistent when projecting multiple related high-dimensional sources.

Core UMAP parameters include n_neighbors, n_components, metric, and min_dist. These control local neighborhood size, output dimensionality, distance measurement, and minimum embedding distance for manifold learning.