|
Abstract:
|
Linear Discriminant Analysis (LDA) is a classification technique used for multivariate data. Each discriminant is a linear combination of all the variables. However, when the number of variables (p) is larger than the sample size (n), all variables cannot be used because the matrices in the discriminant function become singular. There are several methods for overcoming this difficulty: performing a principal component analysis (PCA), and then performing the LDA on the principal components; using univariate significance tests to choose a subset of the variables for the LDA; forward, backward, or stepwise variable selection; or more complex techniques. We compare these techniques in terms of their predictive ability and their ability to select the important features. These comparisons will be performed over several simulated data sets of various complexities and a real dataset.
|
