UMAP-learn Dimensionality Reduction Skill

What it does

Gives Claude a practical playbook for nonlinear dimensionality reduction with umap-learn.

• Deep parameter guide for n_neighbors, min_dist, n_components, and metric, including effects at each value range
• Distinct presets for visualization, clustering preprocessing, and ML feature engineering
• Supervised and semi-supervised UMAP (y=labels, -1 for unlabeled points)
• End-to-end UMAP + HDBSCAN clustering workflow with Adjusted Rand Score evaluation
• Advanced paths: Parametric UMAP with Keras encoders, inverse_transform, AlignedUMAP for time-series/batch data
• Troubleshooting table for fragmented clusters, NaNs, collapsed embeddings, and module shadowing

Who it's for

• Data scientists and computational biologists visualizing high-dimensional data
• Teams replacing t-SNE with something faster and better at global structure
• ML engineers embedding text/image vectors before density-based clustering
• Pipeline authors who must project unseen data into a learned embedding space

Examples

1. "Plot a 2D UMAP of this expression matrix" → standardizes features, fits UMAP(n_neighbors=15, min_dist=0.1, random_state=42), and returns matplotlib scatter code.
2. "Find clusters in this dataset" → switches to clustering settings (n_neighbors=30, min_dist=0.0, n_components=10), runs HDBSCAN, reports cluster counts and noise points.
3. "I need to embed new samples after training" → explains the distribution-consistency caveat of transform() and shows a Parametric UMAP encoder plus save()/load_ParametricUMAP().

Install the skill I found on Claude Skill Mart.
Copy the skills/umap-learn folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/umap-learn/.
When it's done, tell me in one line what this skill can do.