Interpret Components and Loss

Compression trades detail for simplicity

A component's loadings are the weights assigned to original features. Large positive or negative loadings show which features shape that direction, though correlated features and sign conventions can make simple stories unreliable. Component signs can flip without changing the represented geometry.

Reconstruction approximately maps compressed points back to original space. Reconstruction error measures lost detail and can help detect unusual observations, but high error may reflect a valid rare case rather than corruption.

pca = pipe.named_steps['pca']
print(pca.explained_variance_ratio_)
print(pca.components_[0])  # first component loadings

Tip: Reach for PCA when many numeric features are correlated, visualization or speed matters, and reduced interpretability is acceptable. Skip it when original-feature explanations are central or nonlinear structure dominates.

Note: Embeddings later in the track are also dense lower-dimensional representations, but they are learned for semantic objectives rather than produced by PCA's maximum-variance rule.