umap-learn

Dimensionality reduction condenses high-dimensional data into a lower-dimensional embedding for visualization and downstream analysis. This Skill computes compact, human-interpretive representations using a single fit_transform step after standardizing features.

Preprocessing high-dimensional data with UMAP embeddings creates a compact structure that density-based methods like HDBSCAN can easily cluster. This approach helps discover underlying patterns in complex datasets across domains like biology, text, and images.

Supervised and semi-supervised UMAP variants are fully supported for feature engineering. These variants allow you to incorporate label information during dimensionality reduction, producing embeddings optimized for downstream machine learning workflows.

This approach provides fast, scalable dimensionality reduction suitable for large datasets. It applies configurable parameters like n_neighbors, min_dist, n_components, and metric to efficiently process high-dimensional data across various analytical domains.

Parametric UMAP uses neural networks to learn the embedding function, allowing you to transform new data without retraining. It is needed when you want to apply the learned transformation to unseen data or integrate embeddings into existing models.

Aligned UMAP computes embeddings across multiple datasets while preserving relationships between them. It ensures that corresponding points across different views remain proximate, enabling comparative analysis of related high-dimensional data sources.