PCA가 성분을 찾는 방법

PCA starts from one idea: the directions where your data spreads most carry the most information. That spread is called variance.

PCA hunts for the single axis along which points vary the most. That direction becomes the first principal component.

The next axis must be at right angles to the first while capturing the most remaining spread. Each new component is orthogonal to the others.

Together these directions form a fresh coordinate system. Each principal component is a blend of your original features, not just one column.

PCA orders components by how much variance each captures, so the first few hold the bulk of the signal and the last few hold mostly noise.

To shrink your data you project it onto the top components, swapping many original columns for a few informative scores.

The explained variance ratio tells you the fraction of total spread each component keeps, so you can see how much you save.

In scikit-learn you set how many components you want, then fit_transform turns your wide table into compact scores.

After fitting, check explained_variance_ratio_ to see how much each component kept. Summing them shows your total retained variance.

Each component is a weighted sum of features, so PCA only captures linear structure. Curved patterns need other tools.

You trade a little accuracy for far fewer columns. Dropped components hold the least variance, so the loss is usually small.

One rule decides which direction PCA picks first.

PCA finds orthogonal axes of maximum variance, ranks them, and projects data onto the top ones to compress it. 🎯