Umap Learn

If you need to squash high-dimensional data down for visualization or clustering, this is the modern answer. It's faster than t-SNE and actually preserves both local and global structure, which matters when you're trying to understand what your data is doing. The supervised mode is legitimately useful for metric learning, and it scales to higher output dimensions without falling apart, so you can use it as preprocessing for HDBSCAN or other clustering algorithms. Just remember to standardize your features first or you'll get garbage. The parameter guide here is solid: bump n_neighbors to 30 and min_dist to 0.0 for clustering, keep defaults for quick visualization. It follows scikit-learn conventions, so it drops right into existing pipelines.