PCA: 분산, 고유벡터 및 주성분

As feature count grows, datasets become increasingly sparse — the curse of dimensionality. Many features are redundant or correlated, carrying overlapping information. Principal Component Analysis (PCA) solves this by finding a new, smaller set of axes (principal components) that capture the maximum variance in the data with the fewest dimensions.

PCA seeks directions in feature space along which the data varies the most. A direction with high variance captures rich information; a direction with near-zero variance is essentially noise. The first principal component (PC1) is the direction of maximum variance, PC2 is orthogonal to PC1 with the next highest variance, and so on.

PCA operates on the covariance matrix of the centred data. The eigenvectors of this matrix point in the directions of maximum variance, and the corresponding eigenvalues measure how much variance each direction captures. The eigenvectors are the principal components; sorting them by eigenvalue in descending order gives PC1, PC2, ... PCn.

The explained variance ratio of each component is its eigenvalue divided by the sum of all eigenvalues. If PC1 explains 90% of variance and PC2 explains 8%, the first two components together retain 98% of all information. This ratio guides how many components to keep — a common threshold is 95%.

Set n_components as an integer (e.g., PCA(n_components=10)) to keep exactly 10 components, or as a float between 0 and 1 (e.g., PCA(n_components=0.95)) to automatically keep enough components to explain that fraction of variance. The latter is the cleanest approach for pipelines where you want variance-based truncation without knowing the count upfront.

A scree plot shows explained variance ratio (or eigenvalue) on the y-axis and component index on the x-axis. The plot typically shows a steep drop then a flat plateau. The elbow — where the drop becomes gradual — is another heuristic for the number of components to retain, similar to the elbow method in K-Means.

PCA is sensitive to feature scale. A feature measured in thousands will dominate the covariance matrix. Always standardise with StandardScaler before PCA to give each feature unit variance. Centring (zero mean) is essential — PCA implicitly does this, but if you use a Pipeline, the scaler should come first so PCA operates on already-centred, equal-scale features.

Each principal component is a linear combination of the original features — a weighted sum. Inspecting the component loadings (the coefficients) reveals which original features contribute most to each PC. However, components are often not directly interpretable because they mix features together. PCA is primarily a compression tool, not a feature selection tool.

In practice, scikit-learn computes PCA via Singular Value Decomposition (SVD) rather than explicit eigendecomposition of the covariance matrix, because SVD is numerically more stable and works directly on the data matrix without forming the covariance matrix. The result is mathematically identical. For very large datasets, PCA(svd_solver='randomized') uses an approximate randomised SVD for speed.

Important limitations: PCA finds only linear relationships between features. If the meaningful structure in your data lies on a curved manifold (e.g., a Swiss roll), PCA will not discover it effectively — kernel PCA or t-SNE are better alternatives. Also, PCA components are orthogonal by construction, which can be a mismatch if your underlying factors are correlated.

Here is the full workflow: scale, PCA to 2D, and scatter-plot with class colour to check if the reduced space still separates classes visually. This is a standard exploratory step before training a classifier on the full feature set.

In this lesson you learned: PCA finds directions of maximum variance via the covariance matrix eigenvectors, explained variance ratio guides how many components to keep (typically aim for 95%), and always standardise features before PCA so scale differences do not bias the components. Next up we project data into principal-component space and reconstruct it to quantify information loss.