Nonlinear embedding methods for visualization
Both t-SNE (t-Distributed Stochastic Neighbor Embedding) and UMAP (Uniform Manifold Approximation and Projection) are nonlinear dimensionality reduction techniques for visualizing high-dimensional data in 2D or 3D. Unlike PCA, which preserves global variance, these methods emphasize local structure: nearby points in the high-D space stay nearby in the low-D plot, and distant points stay far.
This local-structure emphasis makes clusters visually apparent. You can quickly spot groups in your data and spot outliers. Both methods are designed for exploratory analysis and visualization, not for use as preprocessing before a predictive model.
Speed, stability, and visual artifacts
t-SNE is older and slower, requiring parameter tuning (perplexity is critical and data-dependent). Results can vary between runs, and the algorithm struggles with very large datasets (millions of points). UMAP is faster and more stable, making it practical for large-scale data exploration. UMAP also tends to preserve more global structure than t-SNE.
Both methods can produce misleading visual artifacts. Spurious cluster separations can appear even in random data if tuned poorly. Always validate clusters with domain knowledge or metrics, not visual inspection alone. Neither t-SNE nor UMAP is appropriate for downstream supervised learning; they can distort relationships in ways that harm classifier performance. Use them for exploration, not as input to models.