unsupervised
Principal component analysis (PCA)
Reduce several correlated numeric features to a smaller set of variance directions.
Requirements
- two or more numeric features
- meaningful feature scaling
- complete observations
Avoid when
- the main structure is strongly nonlinear
- components need an automatic causal interpretation
Formula
X = UΣVᵀ
Interpretation
Explained variance shows compression quality; loadings show which features define each component.
Worked example
Reduce several correlated biomarkers to a smaller set of principal components.
Open example datasets