t-SNE

t-SNE

t-Distributed Stochastic Neighbor Embedding for visualizing high-dimensional data in 2D or 3D.

When to use:

• Need 2D/3D visualization
• Want to reveal cluster structure
• Have moderate dataset (<10k samples)
• Don't need to transform new data
• Exploration and presentation

Strengths: Excellent visualizations, reveals clusters beautifully, preserves local structure Weaknesses: Very slow, no inference on new data, sensitive to hyperparameters, different runs give different results, doesn't preserve global structure

Model Parameters

N Components (default: 2, required) Embedding dimensions (typically 2 or 3 for visualization).

• Max: 3 (designed for visualization)

Perplexity (default: 30.0) Balance between local and global structure. Roughly the number of close neighbors.

• Small (5-15): Emphasizes local structure, many small clusters
• Medium (30-50): Balanced (default)
• Large (50-100): More global structure, fewer clusters
• Rule: Should be less than n_samples
• Larger datasets need larger perplexity

Learning Rate (default: 200.0) Step size for gradient descent optimization.

• Too low (<10): Slow convergence, poor results
• Good range (10-1000): Depends on data
• Too high (>1000): Unstable, poor results
• Try: [10, 100, 200, 500, 1000]

Max Iterations (default: 1000) Number of optimization iterations.

• Minimum 250: Very fast but may not converge
• 1000: Standard (default)
• 2000-5000: Better convergence for difficult data

Metric (default: "euclidean") Distance metric for high-dimensional space:

• euclidean: Standard distance (default)
• manhattan: L1 distance
• cosine: Angle similarity
• correlation: Pearson correlation

Random State (default: 42) Seed for reproducibility (t-SNE is stochastic).