umap-learn

Dimensionality reduction projects high-dimensional data into 2D or 3D embeddings while preserving structural relationships for visual exploration. This approach supports various metrics and follows an sklearn-compatible API for efficient implementation.

UMAP serves as effective clustering preprocessing by reducing high-dimensional data to meaningful low-dimensional embeddings before applying HDBSCAN. This workflow improves clustering performance on complex domains like biology, text, and images.

Parametric UMAP supports optional neural network-based embeddings for complex pipelines. It extends standard manifold learning by enabling supervised embeddings and transform operations for labeled data processing.

Manifold learning through this sklearn-compatible implementation supports transform and inverse_transform operations. You can fit the model on training data and apply the learned transformation to project new unseen samples into the existing embedding space.

UMAP provides fast dimensionality reduction specifically designed to preserve both local and global data structure. It handles various distance metrics and supports supervised learning, making it highly adaptable for complex domains like biology and text.

You do not need explicit sklearn installation, but the tool follows an sklearn-compatible API. This design ensures seamless integration into existing machine learning pipelines for generating efficient 2D and 3D embeddings.