rdarob: Robust Regularized Discriminant Analysis
In abnormally-distributed/cvreg: Cross Validation and Robust Estimation Utilities

Description Usage Arguments Value Examples

This model utilizes a combination of shrinkage and robust estimation in order to improve the performance of the standard quadratic discriminant. Robustness is obtained by utilizing the minimum covariance determinant covariance estimate. Regularization is performed in two ways: first each within-level covariance matrix is shrunk towards a pooled covariance matrix. The amount of shrinkage for this is controlled by the hyperparameter α. Second, each covariance matrix is shrunk towards an identity matrix scaled by the geometric mean of the diagonal of a respective matrix. This is controlled by the hyperparameter γ. A guide to help interpret how the two hyperparameters interact is given below.

γ = 0 ; α = 0 : Equivalent to robust qda.
γ = 0 ; α = 1 : Equivalent to robust lda.
γ = 1 ; α = 0 : Levels are conditionally independent and within-level covariances are the scaled diagonal.
γ = 1 ; α = 1 : Classification based on robust distance to the nearest mean. Similar to k-means.

Note that the returned object is of class "qdarob" and hence is compatible with its methods.

rdarob(
  formula,
  data,
  grouping,
  prior = proportions,
  gamma = 0.5,
  alpha = 0.5,
  kappa = 0.75
)

`formula`	a model formula
`data`	a data frame
`prior`	a vector of prior probabilities for each class
`gamma`	within-level covariance shrinkage parameter.
`alpha`	parameter controlling shrinkage towards pooled covariance.
`kappa`	the minimum proportion of observations retained in the minimum covariance determinant estimators of the covariances. defaults to 0.75.