My t-SNE plot shows clear clusters. Can I run K-Means on the 2D coordinates, or feed them into a model?
Answer 1: Avoid using t-SNE output as model features
t-SNE has several problems as a preprocessing step. It can't transform new data (Q8). Its output changes between runs. It distorts distances and densities (Q9). And if you run it on all data before splitting, the test data has influenced the layout, which leaks information. A classifier built on t-SNE coordinates usually looks good in experiments but generalizes poorly. Use PCA or UMAP instead if you need reduced features.
