Each UMAP run gives a different plot. When I set random_state, I get a warning and it becomes much slower.
Why this happens
UMAP uses randomness in its approximate neighbor search and in its optimization. Results differ between runs unless you fix the seed. However, UMAP's parallel optimization can't be made fully deterministic, so setting random_state forces it to run on a single thread. That's what the warning means, and why it's slower.
Answer 1: Set random_state when reproducibility matters
reducer = umap.UMAP(random_state=42)
Accept the slower single-threaded run for final figures, published results, or anything you need to reproduce exactly.
