t-SNE on 100,000 rows with 300 features has been running for hours.
Why this happens
t-SNE compares every point with its neighbors over many iterations. scikit-learn's default Barnes-Hut method is much faster than the exact method, but it still scales poorly with large datasets, and computing neighbors in high dimensions is expensive.
Answer 1: Reduce to about 50 dimensions with PCA first
This is the standard recommendation. PCA removes noise and makes the neighbor computation far cheaper, and the resulting t-SNE plot is usually just as good or better.
from sklearn.pipeline import make_pipeline
X_50 = PCA(n_components=50,
random_state=0).fit_transform(X_scaled)
emb = TSNE(perplexity=30, random_state=42,
n_jobs=-1).fit_transform(X_50)
n_jobs=-1 parallelizes the neighbor search across CPU cores.
