umap-learn

Dimensionality reduction using UMAP transforms high-dimensional data into fast, low-dimensional embeddings suitable for 2D or 3D visualization. You can configure parameters like n_neighbors, min_dist, n_components, and metric to shape the resulting embedding.

UMAP embeddings serve as effective clustering preprocessing, reducing dimensionality to reveal underlying data structure before applying density-based clustering. This workflow integrates smoothly into standard machine learning pipelines to prepare features for HDBSCAN.

Cross-dataset alignment maps related high-dimensional datasets into a shared low-dimensional embedding space using UMAP. This ensures structural consistency across different data sources, allowing accurate comparative visualization and analysis.

Supervised and semi-supervised embeddings are supported, allowing UMAP to utilize label information during the dimensionality reduction process. This guides the embedding to separate labeled classes while maintaining topological structure for unlabeled points.

Parametric UMAP uses neural networks to learn the embedding function, enabling inverse_transform to map low-dimensional points back to the original high-dimensional space. This provides a trained model for embedding new data and generating feature representations.

Adjust n_neighbors to control local versus global structure preservation and min_dist to manage embedding compactness during dimensionality reduction. Tuning these parameters balances cluster tightness against the broader topological layout in the visualization.