umap-learn

Dimensionality reduction embeds high-dimensional datasets into 2D or 3D space for visualization while preserving essential data structure. This approach facilitates visual exploration of complex datasets by mapping manifold relationships into lower dimensions.

Preprocess data for density-based clustering by applying manifold learning to preserve local structure before running algorithms like HDBSCAN. This dimensionality reduction step ensures local neighborhoods remain intact, yielding more accurate density-based cluster assignments.

Yes, supervised and semi-supervised learning can incorporate label information to guide the embedding process for better class separation. By using labels during manifold approximation, the transformed features provide improved discriminative power for downstream models.

UMAP supports feature engineering by embedding high-dimensional data into intermediate dimensions to create transformed features. These engineered features preserve both local and global data structure, enhancing the performance of downstream machine learning models.

Parametric UMAP leverages neural networks to learn the manifold embedding function, allowing extension to new unseen data. This advanced capability supports aligned UMAP for handling time-series or shifting distributions across multiple datasets.

Avoid UMAP when exact global distances must be strictly preserved or when interpretability of individual feature axes is required. As a non-linear manifold learning technique, the transformed axes lack direct semantic meaning, limiting straightforward feature interpretation.