t-SNE & UMAP — High-Dimensional Visualization

import numpy as np
import matplotlib.pyplot as plt
from sklearn.manifold import TSNE
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import load_digits

# DIGITS DATASET: 8x8 images = 64 features, 10 classes (0-9)
digits = load_digits()
X, y = digits.data, digits.target

print(f"Digits dataset: {X.shape} features -> visualizing in 2D")

# SCALE BEFORE TSNE (recommended)
X_sc = StandardScaler().fit_transform(X)

# PCA first to speed up t-SNE (recommended for large datasets)
X_pca50 = PCA(n_components=50, random_state=42).fit_transform(X_sc)

# t-SNE
tsne = TSNE(n_components=2, perplexity=30, n_iter=1000, random_state=42, learning_rate="auto", init="pca")
X_tsne = tsne.fit_transform(X_pca50)

# PCA 2D for comparison
pca = PCA(n_components=2, random_state=42)
X_pca2d = pca.fit_transform(X_sc)

fig, axes = plt.subplots(1, 2, figsize=(16, 7))
colors = plt.cm.tab10(np.linspace(0, 1, 10))

for ax, (X_embed, method) in zip(axes, [(X_pca2d, "PCA"), (X_tsne, "t-SNE")]):
    for digit in range(10):
        mask = y == digit
        ax.scatter(X_embed[mask, 0], X_embed[mask, 1], c=[colors[digit]], label=str(digit),
                   s=20, alpha=0.7)
    ax.set_title(f"{method} Embedding -- Digit Recognition\nEach color = one digit class")
    ax.legend(title="Digit", loc="upper right", fontsize=8, markerscale=1.5)

plt.tight_layout()
plt.savefig("tsne_pca.png", dpi=100, bbox_inches="tight")
plt.show()

# t-SNE PARAMETERS
print("t-SNE critical parameters:")
params = {
    "perplexity":     "Controls balance of local vs global structure. Range: 5-50. Try 30 first.",
    "n_iter":         "More iterations = better convergence. 1000 minimum; use 2000 for production.",
    "learning_rate":  "Use 'auto' (sklearn>=1.2). Manual: n_samples/12 is a common heuristic.",
    "init":           "'pca' gives better results than random initialization. Use 'pca'.",
    "random_state":   "Fix for reproducibility -- t-SNE is stochastic, different runs look different.",
}
for param, tip in params.items():
    print(f"  {param:16s}: {tip}")

print("\nCRITICAL RULES for t-SNE:")
rules = [
    "Do NOT use t-SNE coordinates as features in models -- distances are not globally meaningful",
    "Same class does NOT mean nearby in t-SNE -- cluster sizes are meaningless",
    "Different runs produce different layouts -- fix random_state",
    "Run PCA first (to ~50 dims) before t-SNE when n_features >> 50",
    "UMAP (pip install umap-learn) is faster, preserves global structure better, more reproducible",
]
for rule in rules:
    print(f"  -> {rule}")