umap-learn

Nonlinear dimensionality reduction for high-dimensional data visualization is achieved by applying Uniform Manifold Approximation and Projection (UMAP) to generate 2D or 3D embeddings. This preserves essential data structure while reducing complexity for intuitive scatter plot exploration.

Yes, you can use UMAP to preprocess data for density-based clustering algorithms like HDBSCAN. Reducing dimensions first prepares the high-dimensional data, making subsequent clustering tasks more efficient and structurally coherent.

UMAP supports supervised and semi-supervised learning tasks by incorporating label information to guide the manifold embeddings. This allows the dimensionality reduction to respect known class boundaries during the transformation process.

To run manifold learning with umap-learn, you need the umap-learn Python library alongside dependencies like scikit-learn, matplotlib, hdbscan, numba, and tensorflow. These facilitate efficient computation, plotting, and clustering preprocessing.

Choose UMAP over other dimensionality reduction techniques when you need to efficiently reduce dimensions for complex, high-dimensional datasets while preserving the non-linear structure. It is particularly effective for creating 2D/3D visualizations and clustering preprocessing.