How to Use UMAP (umap-learn.readthedocs.io)

This tutorial walks through using UMAP — a general-purpose manifold-learning and dimensionality-reduction algorithm — as a scikit-learn–compatible tool for visualizing and exploring datasets. It shows how UMAP can be used as a drop-in replacement for t-SNE in sklearn pipelines, making it easy to construct, fit, transform and visualize embeddings with the familiar fit/fit_transform/transform API. The guide demonstrates that UMAP preserves meaningful structure: low-dimensional 2D embeddings separate penguin species (333 samples, 4 measured features → 2D) and cluster handwritten digit classes from the sklearn digits dataset (1,797 samples, 64 → 2D), enabling intuitive visual inspection where pairwise scatter matrices become impractical. Key practical and technical points: preprocess numeric features (the tutorial uses StandardScaler and drops NAs), instantiate reducer = umap.UMAP() and call fit_transform or fit then transform; the embedding lives in reducer.embedding_ and transform returns identical results. Default UMAP settings reduce to n_components=2 and use init='spectral', metric='euclidean', n_neighbors=15 and min_dist=0.1 (see printed params), which balance local versus global structure. Because UMAP follows sklearn conventions, it easily integrates into pipelines and supports reproducible runs via random_state — making it a convenient, interpretable choice for exploratory analysis and downstream modeling in higher-dimensional ML workflows.