선형 판별 분석

Linear discriminant analysis has two closely related uses. As a classifier, it estimates class-specific distributions and assigns an observation to the class with the greatest posterior probability under the model. In its common form, each class is modeled with a Gaussian distribution and classes share a covariance matrix. The shared covariance assumption leads to linear decision boundaries. If class spreads differ substantially, the assumption can be unsuitable; quadratic discriminant analysis relaxes it by allowing class-specific covariance matrices. LDA is also a supervised dimensionality-reduction method. It finds projection directions that make class means far apart relative to within-class variation. For K classes, the discriminant subspace has at most K minus one useful directions, because class-mean differences span at most that many dimensions. The projection can help visualize labeled groups or provide compact inputs to another model, but it is optimized for separation among the classes used during fitting. Principal component analysis has a different objective. PCA finds directions of high overall variance without consulting labels. A direction with large variance may reflect within-class variation rather than class separation. Conversely, LDA may emphasize a direction with modest overall variance if that direction separates the labeled classes. Neither projection should be judged as universally superior; the useful representation depends on the task and downstream evaluation. LDA's assumptions and data conditions matter. Features should be numeric or appropriately encoded, and covariance estimates can be unstable when the number of features is large relative to examples. Shrinkage or dimensionality reduction may help in some cases. Near-duplicate variables and poorly scaled or collinear data can also cause numerical issues depending on the solver. Check whether classes have enough observations to estimate the model. Fit the entire preprocessing and classifier pipeline inside each training fold. Evaluate predictive performance using metrics suitable for class balance and error costs, and inspect calibration separately if probabilities will guide decisions. A visually separated projection alone does not establish reliable generalization.