umap-learn

Dimensionality reduction compresses high-dimensional data into low-dimensional embeddings while preserving structural relationships. You need it when visualizing complex datasets, preprocessing features for clustering, or preparing data for downstream machine learning pipelines.

To reduce high-dimensional data to a 2D embedding, train a UMAP model on your standardized data. You can configure parameters like n_neighbors, min_dist, and metric to generate compact visualizations tailored to your dataset's structure.

Yes, UMAP embeddings work well as a preprocessing step for clustering algorithms like HDBSCAN. By reducing dimensions first, clustering algorithms can operate more efficiently and identify meaningful groups in the lower-dimensional space.

UMAP supports supervised and semi-supervised embedding workflows, allowing you to align related datasets. You can use transform options to project new data into an existing embedding space for consistent downstream analysis.

The n_neighbors parameter controls local versus global structure preservation, while min_dist controls how tightly points are packed in the low-dimensional space. Adjusting these parameters balances fine detail against broader topological patterns in the visualization.

To align embeddings across related datasets, use UMAP's cross-dataset alignment capabilities. This ensures consistent mapping of new data points into an existing low-dimensional space, maintaining structural continuity for comparative analysis.