PCA

PCA

Principal Component Analysis finds orthogonal directions of maximum variance in the data.

When to use:

• Need interpretable linear combinations
• Want to remove correlated features
• Data has linear structure
• Need fast, scalable solution
• First choice for most problems

Strengths: Fast, scalable, interpretable, reversible, works on new data, no hyperparameters Weaknesses: Linear only, sensitive to scaling, assumes Gaussian-like distributions

Model Parameters

N Components (default: 2, required) Number of principal components to keep.

• 2-3: Visualization
• Based on explained variance: Keep components explaining 80-95% variance
• Rule of thumb: min(n_samples, n_features)

SVD Solver (default: "auto") Algorithm to compute singular value decomposition:

• auto: Automatically choose based on data shape (default)
• full: Exact, slow, uses standard LAPACK solver
• arpack: Faster for small n_components, iterative
• randomized: Very fast approximation for large datasets

Whiten (default: false) Transform components to have unit variance.

• false: Components scaled by explained variance (default)
• true: All components have equal variance (useful before clustering/classification)

Random State (default: 42) Seed for reproducibility (used with randomized solver).