PCA reduces dimensionality by rotating to uncorrelated components that maximize variance. It clarifies structure and speeds up downstream modeling.
Preparation
- Standardize features (z‑score) so scale doesn’t dominate
- Handle missing values and remove or cap extreme outliers
- Optionally whiten for unit variance across components
Choosing Components
- Scree plot elbow and cumulative variance thresholds (e.g., 90–95%)
- Domain constraints: interpretability vs. compression
- Cross‑validate downstream model performance with k components
Interpreting Results
- Loadings reveal which features drive each component
- Biplots combine scores and loadings to visualize structure
- Component scores can replace raw features in models
Caveats & Alternatives
- PCA is linear and sensitive to scaling
- For nonlinear manifolds: t‑SNE/UMAP for visualization
- For sparsity/interpretability: Sparse PCA or factor analysis